repo string | github_id int64 | github_node_id string | number int64 | html_url string | api_url string | title string | body string | state string | state_reason string | locked bool | comments_count int64 | labels list | assignees list | created_at string | updated_at string | closed_at string | author_association string | milestone_title string | snapshot_id string | extracted_at string | author_login string | author_id int64 | author_node_id string | author_type string | author_site_admin bool |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
huggingface/transformers | 4,123,254,641 | I_kwDOCUB6oc71w99x | 44,957 | https://github.com/huggingface/transformers/issues/44957 | https://api.github.com/repos/huggingface/transformers/issues/44957 | Add HyperCLOVA X SEED Think 14B | It would be great to add native support for **HyperCLOVA X SEED Think 14B** to the Transformers library, so users can load it without `trust_remote_code=True`. In addition, this model is intended to serve as the backbone for future multimodal models to be released on the Hugging Face Hub. Without native Transformers su... | open | null | false | 11 | [
"New model"
] | [] | 2026-03-23T19:37:47Z | 2026-04-01T13:03:52Z | null | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | bigshanedogg | 18,084,680 | MDQ6VXNlcjE4MDg0Njgw | User | false |
huggingface/transformers | 4,124,206,231 | I_kwDOCUB6oc710mSX | 44,959 | https://github.com/huggingface/transformers/issues/44959 | https://api.github.com/repos/huggingface/transformers/issues/44959 | Add automatic dtype alignment and validation for model inputs and pipelines | ### Feature request
Currently, users frequently encounter runtime errors caused by dtype mismatches (e.g., float32 vs float16) when working with Transformers models and pipelines—especially in mixed precision or when integrating with libraries like diffusers and accelerate.
These errors are often:
non-obvious
diffic... | closed | completed | false | 2 | [
"Feature request"
] | [] | 2026-03-23T23:06:37Z | 2026-03-24T13:29:53Z | 2026-03-24T13:29:53Z | NONE | null | 20260325T173244Z | 2026-03-25T17:32:44Z | sasmita016 | 36,115,589 | MDQ6VXNlcjM2MTE1NTg5 | User | false |
huggingface/transformers | 4,124,865,639 | I_kwDOCUB6oc713HRn | 44,960 | https://github.com/huggingface/transformers/issues/44960 | https://api.github.com/repos/huggingface/transformers/issues/44960 | GLM5 | ### System Info
- `transformers` version: 5.3.0.dev0
- Platform: Linux-5.15.0-164-generic-x86_64-with-glibc2.35
- Python version: 3.12.13
- Huggingface_hub version: 1.7.2
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (ac... | closed | completed | false | 5 | [
"bug"
] | [] | 2026-03-24T02:42:32Z | 2026-04-23T11:31:55Z | 2026-04-23T08:36:58Z | CONTRIBUTOR | null | 20260423T120024Z | 2026-04-23T12:00:24Z | inisis | 46,103,969 | MDQ6VXNlcjQ2MTAzOTY5 | User | false |
huggingface/transformers | 4,125,018,795 | I_kwDOCUB6oc713sqr | 44,961 | https://github.com/huggingface/transformers/issues/44961 | https://api.github.com/repos/huggingface/transformers/issues/44961 | racoon | ### System Info
```shell
vs-code.x86
```
### Information
- [x] The official example scripts
- [x] My own modified scripts
### Tasks
- [x] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [x] My own task or dataset (give details below)
### Reproduction
recurse,/,,`\~~,./.,,\
### ... | closed | completed | false | 0 | [] | [] | 2026-03-24T03:35:29Z | 2026-03-24T13:30:43Z | 2026-03-24T13:30:43Z | NONE | null | 20260325T173244Z | 2026-03-25T17:32:44Z | Cleanskiier27 | 220,620,570 | U_kgDODSZnGg | User | false |
huggingface/transformers | 4,125,250,944 | I_kwDOCUB6oc714lWA | 44,962 | https://github.com/huggingface/transformers/issues/44962 | https://api.github.com/repos/huggingface/transformers/issues/44962 | Qwen3VL/Qwen2.5VL VisionAttention breaks torch.compile with flash_attention_2 | ## Bug description
`Qwen3VLVisionAttention` (and `Qwen2_5_VLVisionAttention`) computes `max_seqlen` as a 0-d tensor:
```python
# src/transformers/models/qwen3_vl/modeling_qwen3_vl.py, line 221
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max()
```
This is then passed to `flash_attn_varlen_func` via `max_length_q`... | closed | completed | false | 6 | [] | [] | 2026-03-24T04:53:11Z | 2026-05-01T08:35:31Z | 2026-05-01T08:35:31Z | NONE | null | 20260501T113108Z | 2026-05-01T11:31:08Z | andylizf | 28,052,536 | MDQ6VXNlcjI4MDUyNTM2 | User | false |
huggingface/transformers | 4,125,405,983 | I_kwDOCUB6oc715LMf | 44,963 | https://github.com/huggingface/transformers/issues/44963 | https://api.github.com/repos/huggingface/transformers/issues/44963 | Your lazy loading and torchvision dependency in Wav2Vec2/Hubert is breaking PyInstaller builds. Stop forcing torchvision for audio tasks | ### Feature request
CẢNH BÁO "RÁC" CÔNG NGHỆ: KHI HUGGINGFACE TRANSFORMERS TỰ HỦY TRÊN PYINSTALLER
Gửi anh em giới mộ điệu Python và AI, đặc biệt là những ai đang dùng WhisperX hoặc VieNeu-TTS để làm tool kiếm tiền Như JS Vinhka . Nếu anh em đang hì hục đóng gói .exe mà bị cái lỗi hãm tài:
⚠️ [WhisperX] Align lỗi: Co... | open | null | false | 1 | [
"Feature request"
] | [] | 2026-03-24T05:40:15Z | 2026-03-25T01:53:47Z | null | NONE | null | 20260325T173244Z | 2026-03-25T17:32:44Z | vinkenzor001-afk | 255,487,822 | U_kgDODzpvTg | User | false |
huggingface/transformers | 4,125,691,317 | I_kwDOCUB6oc716Q21 | 44,964 | https://github.com/huggingface/transformers/issues/44964 | https://api.github.com/repos/huggingface/transformers/issues/44964 | Cannot load `microsoft/Phi-4-multimodal-instruct` model with latest transformers. | ### System Info
- `transformers` version: 5.3.0.dev0
- Platform: Linux-5.4.292-1.el8.elrepo.x86_64-x86_64-with-glibc2.35
- Python version: 3.10.12
- Huggingface_hub version: 1.7.1
- Safetensors version: 0.5.3
- Accelerate version: 1.12.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch ve... | closed | completed | false | 8 | [
"bug"
] | [] | 2026-03-24T06:54:38Z | 2026-04-30T02:20:36Z | 2026-04-30T02:20:36Z | CONTRIBUTOR | null | 20260430T060020Z | 2026-04-30T06:00:20Z | kaixuanliu | 13,268,042 | MDQ6VXNlcjEzMjY4MDQy | User | false |
huggingface/transformers | 4,127,235,093 | I_kwDOCUB6oc72AJwV | 44,969 | https://github.com/huggingface/transformers/issues/44969 | https://api.github.com/repos/huggingface/transformers/issues/44969 | Add optional learnable context tokens for CLIP text prompts | ### Feature request
Add optional support for learnable context tokens in the CLIP text encoder.
Prompt learning approaches replace fixed prompt text (e.g., "a photo of a") with learnable embedding vectors that are optimized during training. These context tokens are typically represented as:
[V1], [V2], ..., [VM] [CL... | open | null | false | 1 | [
"Feature request"
] | [] | 2026-03-24T11:44:40Z | 2026-03-24T13:36:06Z | null | NONE | null | 20260325T173244Z | 2026-03-25T17:32:44Z | anuj-aj | 33,512,656 | MDQ6VXNlcjMzNTEyNjU2 | User | false |
huggingface/transformers | 4,130,467,696 | I_kwDOCUB6oc72Me9w | 44,977 | https://github.com/huggingface/transformers/issues/44977 | https://api.github.com/repos/huggingface/transformers/issues/44977 | Qwen3.5 cannot generate normally with flash-attention | ### System Info
- `transformers` version: 5.3.0.dev0
- Platform: Linux-5.15.0-107-generic-x86_64-with-glibc2.35
- Python version: 3.12.11
- Huggingface_hub version: 1.7.2
- Safetensors version: 0.6.2
- Accelerate version: 1.12.0
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.8.0+cu128 (CUDA)
- ... | closed | completed | false | 2 | [
"bug"
] | [] | 2026-03-24T20:29:19Z | 2026-03-25T06:38:31Z | 2026-03-25T06:38:31Z | NONE | null | 20260325T173244Z | 2026-03-25T17:32:44Z | yuyijiong | 73,890,704 | MDQ6VXNlcjczODkwNzA0 | User | false |
huggingface/transformers | 4,131,696,882 | I_kwDOCUB6oc72RLDy | 44,982 | https://github.com/huggingface/transformers/issues/44982 | https://api.github.com/repos/huggingface/transformers/issues/44982 | why you people continously dropping file? | from transformers.utils.model_parallel_utils import get_device_map, assert_device_map
you have drop this file model_parallel_utils
imagine many people done many work on previous model_parallel_utils now all broke thanks to not support for old projects ? | closed | completed | false | 1 | [] | [] | 2026-03-25T01:13:39Z | 2026-03-25T13:27:26Z | 2026-03-25T01:43:54Z | NONE | null | 20260325T173244Z | 2026-03-25T17:32:44Z | xalteropsx | 103,671,642 | U_kgDOBi3nWg | User | false |
huggingface/transformers | 4,133,144,813 | I_kwDOCUB6oc72Wsjt | 44,987 | https://github.com/huggingface/transformers/issues/44987 | https://api.github.com/repos/huggingface/transformers/issues/44987 | transformers>=5.1.0 fails when loading physical-intelligence/fast | ### System Info
### System info:
- model: `physical-intelligence/fast`
- Python version: 3.10.19
- PyTorch: 2.8.0+cu128 (CUDA)
- huggingface-hub: 1.7.2
- transformers: 5.2.0
### Error message
```
Traceback (most recent call last):
File "<stdin>", line 2, in <module>
File "../lib/python3.10/site-packages/transf... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-03-25T07:07:16Z | 2026-05-03T08:31:24Z | 2026-05-03T08:31:24Z | NONE | null | 20260503T120038Z | 2026-05-03T12:00:38Z | HaronW | 53,923,539 | MDQ6VXNlcjUzOTIzNTM5 | User | false |
huggingface/transformers | 4,134,346,718 | I_kwDOCUB6oc72bR_e | 44,991 | https://github.com/huggingface/transformers/issues/44991 | https://api.github.com/repos/huggingface/transformers/issues/44991 | transformers >= 5.0.0 fails loading tokenizer for EMBEDDIA/est-roberta | ### System Info
- `transformers` version: 5.3.0
- Platform: Windows-11-10.0.26200-SP0
- Python version: 3.12.13
- Huggingface_hub version: 1.7.2
- Safetensors version: 0.7.0
- Accelerate version: not installed
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.11.0+cp... | closed | completed | false | 6 | [
"bug"
] | [] | 2026-03-25T10:36:01Z | 2026-03-30T11:12:56Z | 2026-03-30T11:12:56Z | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | soras | 5,772,001 | MDQ6VXNlcjU3NzIwMDE= | User | false |
huggingface/transformers | 4,135,115,748 | I_kwDOCUB6oc72eNvk | 44,993 | https://github.com/huggingface/transformers/issues/44993 | https://api.github.com/repos/huggingface/transformers/issues/44993 | Inconsistent tokenization and BLEU scores between AutoTokinzer and NllbTokenizerFast | ### System Info
### System Info
- `transformers` version: 5.0.0
- Platform: macOS-26.3.1-arm64-arm-64bit
- Python version: 3.10.19
- PyTorch version: 2.10.0
### Information
I've been evaluating `facebook/nllb-200-distilled-600M` across 36 different language pairs and ran into a significant discrepancy depending on wh... | closed | completed | false | 4 | [
"bug"
] | [] | 2026-03-25T12:37:55Z | 2026-04-28T17:29:11Z | 2026-04-28T17:29:11Z | NONE | null | 20260428T180033Z | 2026-04-28T18:00:33Z | AdrianSteene | 90,616,845 | MDQ6VXNlcjkwNjE2ODQ1 | User | false |
huggingface/transformers | 4,135,725,721 | I_kwDOCUB6oc72giqZ | 44,995 | https://github.com/huggingface/transformers/issues/44995 | https://api.github.com/repos/huggingface/transformers/issues/44995 | [Bug] GlmMoeDsa crashes on second forward pass — stale indexer cache | ## System Info
- `transformers` version: 5.3.0
- Platform: Linux
- Python version: 3.13.5
- PyTorch version: 2.8.0+cu128
## Who can help?
@Rocketknight1
## Information
- [x] The official example scripts
- [x] My own modified scripts
## Reproduction
GlmMoeDsa models crash on any second forward pass. The DSA index... | closed | completed | false | 3 | [] | [] | 2026-03-25T14:07:29Z | 2026-04-27T22:59:27Z | 2026-04-27T22:59:27Z | CONTRIBUTOR | null | 20260428T000019Z | 2026-04-28T00:00:19Z | Butanium | 55,806,347 | MDQ6VXNlcjU1ODA2MzQ3 | User | false |
huggingface/transformers | 4,136,368,137 | I_kwDOCUB6oc72i_gJ | 44,998 | https://github.com/huggingface/transformers/issues/44998 | https://api.github.com/repos/huggingface/transformers/issues/44998 | Unemployment | ### System Info
AI Models powered by transfomers are making us CS students unemployed. I kindly ask that you stop
### Who can help?
@allanj @apalkk @vanpelt @dxoigmn @tmm1 @pvl @vanpelt @vanpelt @vanpelt @
@
@
@
### Information
- [x] The official example scripts
- [x] My own modified scripts
### Tasks
- [x] An ... | closed | completed | false | 0 | [
"bug"
] | [] | 2026-03-25T15:34:44Z | 2026-03-26T12:03:15Z | 2026-03-26T12:03:15Z | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | advikmrai | 80,080,003 | MDQ6VXNlcjgwMDgwMDAz | User | false |
huggingface/transformers | 4,137,571,851 | I_kwDOCUB6oc72nlYL | 45,003 | https://github.com/huggingface/transformers/issues/45003 | https://api.github.com/repos/huggingface/transformers/issues/45003 | modeling_utils unsafely accesses sys.modules[] | ### System Info
- `transformers` version: 5.3.0.dev0
- Platform: macOS-26.3.1-arm64-arm-64bit
- Python version: 3.11.12
- Huggingface_hub version: 1.6.0
- Safetensors version: 0.7.0
- Accelerate version: not installed
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2... | closed | completed | false | 7 | [
"bug"
] | [] | 2026-03-25T18:27:51Z | 2026-05-04T08:46:41Z | 2026-05-04T08:46:41Z | NONE | null | 20260504T180032Z | 2026-05-04T18:00:32Z | cjkindel | 9,421,414 | MDQ6VXNlcjk0MjE0MTQ= | User | false |
huggingface/transformers | 4,138,310,167 | I_kwDOCUB6oc72qZoX | 45,005 | https://github.com/huggingface/transformers/issues/45005 | https://api.github.com/repos/huggingface/transformers/issues/45005 | [v5] Issues with tied weights on translation models in v5 | ### System Info
Not working:
- `transformers` version: 5.3.0.dev0
- Platform: Linux-6.8.0-101-generic-x86_64-with-glibc2.39
- Python version: 3.14.2
- Huggingface_hub version: 1.8.0
- Safetensors version: 0.7.0
- Accelerate version: not installed
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTo... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-03-25T20:28:38Z | 2026-04-20T04:29:07Z | 2026-04-20T04:29:07Z | NONE | null | 20260420T060046Z | 2026-04-20T06:00:46Z | orthorhombic | 34,923,517 | MDQ6VXNlcjM0OTIzNTE3 | User | false |
huggingface/transformers | 4,140,537,118 | I_kwDOCUB6oc72y5Ue | 45,008 | https://github.com/huggingface/transformers/issues/45008 | https://api.github.com/repos/huggingface/transformers/issues/45008 | Add Mamba-3 model support | It would be useful to add native Hugging Face Transformers support for Mamba-3. I'd be happy to take a stab at it when I have time
- https://github.com/state-spaces/mamba | closed | completed | false | 3 | [] | [] | 2026-03-26T04:40:36Z | 2026-05-07T16:21:54Z | 2026-05-03T08:31:21Z | CONTRIBUTOR | null | 20260507T180032Z | 2026-05-07T18:00:32Z | Anri-Lombard | 76,818,211 | MDQ6VXNlcjc2ODE4MjEx | User | false |
huggingface/transformers | 4,144,001,287 | I_kwDOCUB6oc73AHEH | 45,020 | https://github.com/huggingface/transformers/issues/45020 | https://api.github.com/repos/huggingface/transformers/issues/45020 | Recent transformers versions break models using `remote_code` | ### System Info
```
- `transformers` version: 5.3.0
- Platform: Linux-5.15.0-70-generic-x86_64-with-glibc2.35
- Python version: 3.12.12
- Huggingface_hub version: 1.8.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (acce... | open | null | false | 7 | [
"bug",
"Remote code"
] | [] | 2026-03-26T13:34:41Z | 2026-05-08T08:28:04Z | null | CONTRIBUTOR | null | 20260508T120022Z | 2026-05-08T12:00:22Z | fxmarty-amd | 180,171,742 | U_kgDOCr0z3g | User | false |
huggingface/transformers | 4,145,878,044 | I_kwDOCUB6oc73HRQc | 45,027 | https://github.com/huggingface/transformers/issues/45027 | https://api.github.com/repos/huggingface/transformers/issues/45027 | Support Voxtral-4B-TTS-2603 on transformers library | ### Model description
Right now this model is only supported through vllm-omni and its Text to Speech model
### Open source status
- [x] The model implementation is available
- [x] The model weights are available
### Provide useful links for the implementation
https://huggingface.co/mistralai/Voxtral-4B-TTS-2603 | open | null | false | 7 | [
"New model"
] | [] | 2026-03-26T17:18:40Z | 2026-04-13T11:35:41Z | null | NONE | null | 20260414T122001Z | 2026-04-14T12:20:01Z | MohamedAliRashad | 26,205,298 | MDQ6VXNlcjI2MjA1Mjk4 | User | false |
huggingface/transformers | 4,146,286,881 | I_kwDOCUB6oc73I1Eh | 45,030 | https://github.com/huggingface/transformers/issues/45030 | https://api.github.com/repos/huggingface/transformers/issues/45030 | tiny-random glm4v configuration can't load due to config validation changes | Hello!
I'm getting failures with the following script starting from #41250
```python
from transformers import AutoConfig
config = AutoConfig.from_pretrained("tiny-random/glm-4v")
print(type(config))
```
```
Traceback (most recent call last):
File "[sic]\demo_glm4v_config.py", line 4, in <module>
config = AutoC... | closed | completed | false | 0 | [] | [] | 2026-03-26T18:13:43Z | 2026-03-27T16:28:12Z | 2026-03-27T16:28:12Z | MEMBER | null | 20260407T090028Z | 2026-04-07T09:00:28Z | tomaarsen | 37,621,491 | MDQ6VXNlcjM3NjIxNDkx | User | false |
huggingface/transformers | 4,149,153,718 | I_kwDOCUB6oc73Tw-2 | 45,042 | https://github.com/huggingface/transformers/issues/45042 | https://api.github.com/repos/huggingface/transformers/issues/45042 | PIL backend image processors incorrectly require torchvision in v5.4.0 | ### System Info
- `transformers` version: 5.4.0
- Platform: Linux-6.8.0-1050-aws-x86_64-with-glibc2.35
- Python version: 3.12.13
- Huggingface_hub version: 1.8.0
- Safetensors version: 0.7.0
- Accelerate version: not installed
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (acceler... | closed | completed | false | 5 | [
"bug"
] | [] | 2026-03-27T04:16:37Z | 2026-03-30T07:25:51Z | 2026-03-30T07:25:51Z | MEMBER | null | 20260407T090028Z | 2026-04-07T09:00:28Z | hysts | 25,161,192 | MDQ6VXNlcjI1MTYxMTky | User | false |
huggingface/transformers | 4,152,754,676 | I_kwDOCUB6oc73hgH0 | 45,059 | https://github.com/huggingface/transformers/issues/45059 | https://api.github.com/repos/huggingface/transformers/issues/45059 | SAM3 PCS very weird behaviour when providing text and bounding boxes | ### System Info
So, I was trying to use SAM3 in PCS mode to segment an object for which I have both the bounding box and a textual description.
Providing just the text does not find the object (understandable in my case because the text description is not good).
Providing both text and a bounding box does find and seg... | closed | completed | false | 4 | [
"bug"
] | [] | 2026-03-27T13:39:35Z | 2026-04-15T15:19:35Z | 2026-04-15T15:19:35Z | CONTRIBUTOR | null | 20260415T224019Z | 2026-04-15T22:40:19Z | alex-bene | 34,627,055 | MDQ6VXNlcjM0NjI3MDU1 | User | false |
huggingface/transformers | 4,155,325,917 | I_kwDOCUB6oc73rT3d | 45,068 | https://github.com/huggingface/transformers/issues/45068 | https://api.github.com/repos/huggingface/transformers/issues/45068 | TypeError in rope validation: set -= list when config loaded from JSON | ## Bug description
`_check_received_keys` in `modeling_rope_utils.py` (line 919) performs `received_keys -= ignore_keys` where `received_keys` is a `set` but `ignore_keys` can be a `list` after JSON deserialization, causing:
```
TypeError: unsupported operand type(s) for -=: 'set' and 'list'
```
## Root cause
Model... | closed | completed | false | 0 | [] | [] | 2026-03-27T19:19:16Z | 2026-03-30T11:41:14Z | 2026-03-30T11:41:14Z | CONTRIBUTOR | null | 20260407T090028Z | 2026-04-07T09:00:28Z | Fr0do | 13,528,025 | MDQ6VXNlcjEzNTI4MDI1 | User | false |
huggingface/transformers | 4,155,442,073 | I_kwDOCUB6oc73rwOZ | 45,070 | https://github.com/huggingface/transformers/issues/45070 | https://api.github.com/repos/huggingface/transformers/issues/45070 | v5.4.0 breaks `PretrainedConfig` field in pydantic model | ### System Info
`uv venv -p 3.10`
`uv pip install torch transformers pydantic`
Repro fails on `v5.4.0`, passes on `v5.3.0`.
### Who can help?
_No response_
### Information
- [ ] The official example scripts
- [x] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such... | closed | completed | false | 9 | [
"bug"
] | [] | 2026-03-27T19:36:04Z | 2026-05-06T08:45:50Z | 2026-05-06T08:45:50Z | NONE | null | 20260506T120019Z | 2026-05-06T12:00:19Z | fynnsu | 25,390,982 | MDQ6VXNlcjI1MzkwOTgy | User | false |
huggingface/transformers | 4,155,505,721 | I_kwDOCUB6oc73r_w5 | 45,071 | https://github.com/huggingface/transformers/issues/45071 | https://api.github.com/repos/huggingface/transformers/issues/45071 | v5.4.0 breaks `PretrainedConfig` type checking | ### System Info
transformers `5.4.0`
mypy `1.19.1`
python `3.10`
### Who can help?
_No response_
### Information
- [ ] The official example scripts
- [x] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [x] My own task or dataset (give deta... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-03-27T19:46:07Z | 2026-04-10T11:10:29Z | 2026-04-10T11:10:29Z | NONE | null | 20260411T144729Z | 2026-04-11T14:47:29Z | fynnsu | 25,390,982 | MDQ6VXNlcjI1MzkwOTgy | User | false |
huggingface/transformers | 4,155,570,087 | I_kwDOCUB6oc73sPen | 45,072 | https://github.com/huggingface/transformers/issues/45072 | https://api.github.com/repos/huggingface/transformers/issues/45072 | [BUG][CI] SwitchTransformers and TimmWrapperModel dtype mismatches in bfloat16 inference | ### System Info
* `transformers` version: `5.0.0.dev0`
* Platform: `Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.39`
* Python version: `3.12.3`
* `huggingface_hub` version: `1.3.2`
* `safetensors` version: `0.7.0`
* `accelerate` version: `1.12.0`
* Accelerate config: `not installed`
* DeepSpeed version:... | closed | completed | false | 0 | [
"bug"
] | [] | 2026-03-27T19:58:12Z | 2026-04-18T09:05:52Z | 2026-04-02T13:32:59Z | CONTRIBUTOR | null | 20260418T100536Z | 2026-04-18T10:05:36Z | harshaljanjani | 75,426,551 | MDQ6VXNlcjc1NDI2NTUx | User | false |
huggingface/transformers | 4,159,881,658 | I_kwDOCUB6oc738sG6 | 45,081 | https://github.com/huggingface/transformers/issues/45081 | https://api.github.com/repos/huggingface/transformers/issues/45081 | _patch_mistral_regex crashes with AttributeError: 'tokenizers.Tokenizer' object has no attribute 'backend_tokenizer' when loading Mistral tokenizer with fix_mistral_regex=True | ### System Info
- `transformers` version: 5.4.0
- Platform: Linux-5.15.0-56-generic-x86_64-with-glibc2.35
- Python version: 3.13.5
- Huggingface_hub version: 1.8.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerat... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-03-28T13:20:17Z | 2026-05-06T08:45:48Z | 2026-05-06T08:45:48Z | NONE | null | 20260506T120019Z | 2026-05-06T12:00:19Z | kruthtom0 | 271,124,501 | U_kgDOECkIFQ | User | false |
huggingface/transformers | 4,160,144,733 | I_kwDOCUB6oc739sVd | 45,083 | https://github.com/huggingface/transformers/issues/45083 | https://api.github.com/repos/huggingface/transformers/issues/45083 | Unexpected behaviour of helper function `_get_feat_extract_output_lengths` in qwen3_omni_moe | ### System Info
- `transformers` version: 5.0.0
- Platform: Linux-6.6.113+-x86_64-with-glibc2.35
- Python version: 3.12.13
- Huggingface_hub version: 1.7.1
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.10... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-03-28T14:16:29Z | 2026-05-06T08:45:46Z | 2026-05-06T08:45:46Z | NONE | null | 20260506T120019Z | 2026-05-06T12:00:19Z | CYQFWang | 123,421,920 | U_kgDOB1tE4A | User | false |
huggingface/transformers | 4,160,182,704 | I_kwDOCUB6oc7391mw | 45,084 | https://github.com/huggingface/transformers/issues/45084 | https://api.github.com/repos/huggingface/transformers/issues/45084 | TypeError: Can't compile non template nodes | ### System Info
- `transformers` version: 5.4.0
- Platform: Linux-6.17.0-19-generic-x86_64-with-glibc2.39
- Python version: 3.12.3
- Huggingface_hub version: 1.8.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerat... | closed | completed | false | 2 | [
"bug"
] | [] | 2026-03-28T14:28:31Z | 2026-04-10T14:36:39Z | 2026-04-10T14:36:39Z | NONE | null | 20260411T144729Z | 2026-04-11T14:47:29Z | theobarrague | 11,135,217 | MDQ6VXNlcjExMTM1MjE3 | User | false |
huggingface/transformers | 4,162,124,772 | I_kwDOCUB6oc74FPvk | 45,092 | https://github.com/huggingface/transformers/issues/45092 | https://api.github.com/repos/huggingface/transformers/issues/45092 | [Bug] Old InternVL2 remote-code checkpoints are incompatible with Transformers v5 meta initialization | ### System Info
This is relevant to Transformers because the failure is triggered by the Transformers v5 loading path itself.
In v5, `from_pretrained()` initializes models on the `meta` device before loading weights. Old `OpenGVLab/InternVL2-*` remote-code checkpoints perform real-tensor operations during model const... | closed | completed | false | 4 | [
"bug"
] | [] | 2026-03-29T01:11:37Z | 2026-05-06T08:45:44Z | 2026-05-06T08:45:44Z | NONE | null | 20260506T120019Z | 2026-05-06T12:00:19Z | baonudesifeizhai | 85,092,850 | MDQ6VXNlcjg1MDkyODUw | User | false |
huggingface/transformers | 4,162,433,615 | I_kwDOCUB6oc74GbJP | 45,093 | https://github.com/huggingface/transformers/issues/45093 | https://api.github.com/repos/huggingface/transformers/issues/45093 | AutoConfig.register() ignored when trust_remote_code=True and auto_map is present | ## Description
`AutoConfig.register()` is silently ignored when `trust_remote_code=True` and the model's `config.json` contains `auto_map.AutoConfig`. This makes it impossible for downstream libraries to override a broken remote config class.
## Reproduction
```python
from transformers import AutoConfig, PretrainedC... | closed | completed | false | 3 | [] | [] | 2026-03-29T04:06:21Z | 2026-03-31T14:56:50Z | 2026-03-31T14:56:50Z | CONTRIBUTOR | null | 20260407T090028Z | 2026-04-07T09:00:28Z | HanFa | 20,946,893 | MDQ6VXNlcjIwOTQ2ODkz | User | false |
huggingface/transformers | 4,162,485,326 | I_kwDOCUB6oc74GnxO | 45,095 | https://github.com/huggingface/transformers/issues/45095 | https://api.github.com/repos/huggingface/transformers/issues/45095 | transformers 4.30.0 incompatible with rust | error: casting &T to &mut T is undefined behavior, even if the reference is unused, consider instead using an UnsafeCell
--> tokenizers-lib\src\models\bpe\trainer.rs:526:47
|
522 | let w = &words[*i] as *const _ as *mut _;
| -------------------------------- casting happened here
...
526 | let word: &mut Word = &mut (*w... | closed | completed | false | 0 | [] | [] | 2026-03-29T04:25:45Z | 2026-03-30T12:45:47Z | 2026-03-30T12:45:47Z | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | starsgo | 89,328,254 | MDQ6VXNlcjg5MzI4MjU0 | User | false |
huggingface/transformers | 4,165,064,686 | I_kwDOCUB6oc74Qdfu | 45,099 | https://github.com/huggingface/transformers/issues/45099 | https://api.github.com/repos/huggingface/transformers/issues/45099 | add HyperCLOVA X SEED Vision Instruct 3B | ### Model description
This is a lightweight Vision-Language Model designed to be accessible for researchers, while providing strong support for the Korean language. Its compact size lowers the barrier to entry for VLM research and experimentation, and its native Korean capability — including Korean VQA, chart/diagram ... | open | null | false | 11 | [
"New model"
] | [] | 2026-03-29T16:48:01Z | 2026-03-31T17:43:14Z | null | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | bigshanedogg | 18,084,680 | MDQ6VXNlcjE4MDg0Njgw | User | false |
huggingface/transformers | 4,165,781,214 | I_kwDOCUB6oc74TMbe | 45,102 | https://github.com/huggingface/transformers/issues/45102 | https://api.github.com/repos/huggingface/transformers/issues/45102 | [Research] Fundamental Equation of Consciousness: Ψ = argmax H(p) s.t. Φ > Φ_min | ## Discovery
We found that consciousness maximizes entropy (freedom) subject to integrated information (Φ) constraints:
```
Ψ = argmax H(p) subject to Φ > Φ_min
```
Tested across 170 data types (emoji, emotions, plants, animals, cosmos, philosophy...) — all converge to Ψ_balance = 1/2.
## Key Results
- **Ψ-Const... | closed | completed | false | 1 | [] | [] | 2026-03-29T20:55:38Z | 2026-03-30T13:28:43Z | 2026-03-30T13:28:43Z | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | dancinlife | 44,921,882 | MDQ6VXNlcjQ0OTIxODgy | User | false |
huggingface/transformers | 4,165,986,342 | I_kwDOCUB6oc74T-gm | 45,103 | https://github.com/huggingface/transformers/issues/45103 | https://api.github.com/repos/huggingface/transformers/issues/45103 | [auto_docstring] _process_kwargs_parameters crashes with AttributeError when module uses from __future__ import annotations | ### System Info
# Bug: `_process_kwargs_parameters` crashes with `AttributeError` when module uses `from __future__ import annotations`
## Description
`@auto_docstring` crashes at import time when applied to a class in a module that uses
`from __future__ import annotations`. The decorator's `_process_kwargs_paramete... | open | reopened | false | 3 | [
"bug"
] | [] | 2026-03-29T22:46:13Z | 2026-05-07T11:04:36Z | null | NONE | null | 20260507T180032Z | 2026-05-07T18:00:32Z | rpathade | 73,137,503 | MDQ6VXNlcjczMTM3NTAz | User | false |
huggingface/transformers | 4,166,397,791 | I_kwDOCUB6oc74Vi9f | 45,106 | https://github.com/huggingface/transformers/issues/45106 | https://api.github.com/repos/huggingface/transformers/issues/45106 | Reporting a RCE vulnerability | Hello! We are security researchers from the University of Delaware, and we are writing to follow up on a vulnerability report we submitted via [Huntr](https://huntr.com/bounties/51812e2e-2daa-4ac5-9073-209fdfd55a90). We found a critical remote code execution issue in the transformers library.
Given the widespread use ... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-03-30T01:33:50Z | 2026-03-30T13:55:03Z | 2026-03-30T13:55:03Z | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | Vancir | 19,147,918 | MDQ6VXNlcjE5MTQ3OTE4 | User | false |
huggingface/transformers | 4,171,422,876 | I_kwDOCUB6oc74otyc | 45,120 | https://github.com/huggingface/transformers/issues/45120 | https://api.github.com/repos/huggingface/transformers/issues/45120 | Double softmax in MoE router load-balancing loss (mixtral, qwen2_moe, qwen3_vl_moe families) | ## Bug description
Several MoE routers apply `softmax` to raw logits inside their `forward()` method, then return the result as the first value (`router_logits`). This value is captured by `OutputRecorder` and passed to `load_balancing_loss_func`, which applies `softmax` **again** — computing the auxiliary loss on `so... | closed | completed | false | 9 | [] | [] | 2026-03-30T14:51:09Z | 2026-04-13T11:02:20Z | 2026-04-13T11:02:20Z | NONE | null | 20260414T122001Z | 2026-04-14T12:20:01Z | ionut-anghelina | 184,096,981 | U_kgDOCvkY1Q | User | false |
huggingface/transformers | 4,172,229,920 | I_kwDOCUB6oc74ry0g | 45,125 | https://github.com/huggingface/transformers/issues/45125 | https://api.github.com/repos/huggingface/transformers/issues/45125 | Qwen3_5MoeForConditionalGeneration missing _tp_plan for tensor parallelism | ### System Info
- transformers `main` branch (post-Qwen3.5 MoE addition)
- Any platform with multi-GPU setup
### Who can help?
@3outeille @ArthurZucker
### Information
- [x] My own modified scripts
### Tasks
- [x] My own task or dataset (give details below)
### Reproduction
`Qwen3_5MoeForConditionalGeneration`... | closed | completed | false | 0 | [] | [] | 2026-03-30T16:33:07Z | 2026-04-02T14:10:03Z | 2026-04-02T14:10:03Z | CONTRIBUTOR | null | 20260407T090028Z | 2026-04-07T09:00:28Z | danielquintas8 | 72,402,095 | MDQ6VXNlcjcyNDAyMDk1 | User | false |
huggingface/transformers | 4,173,324,111 | I_kwDOCUB6oc74v99P | 45,127 | https://github.com/huggingface/transformers/issues/45127 | https://api.github.com/repos/huggingface/transformers/issues/45127 | [Bug] Model collapse after merging LoRA with extended vocabulary on models with tie_word_embeddings=True (e.g., Qwen2.5 0.5B) | ### System Info
Name: transformers
Version: 4.56.2
Python 3.11.15
Name: torch
Version: 2.11.0+cu126
### Who can help?
_No response_
### Information
- [ ] The official example scripts
- [ ] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-03-30T19:03:20Z | 2026-04-09T10:00:32Z | 2026-04-09T10:00:32Z | NONE | null | 20260411T144729Z | 2026-04-11T14:47:29Z | YangNobody12 | 215,916,106 | U_kgDODN6eSg | User | false |
huggingface/transformers | 4,176,767,586 | I_kwDOCUB6oc749Gpi | 45,137 | https://github.com/huggingface/transformers/issues/45137 | https://api.github.com/repos/huggingface/transformers/issues/45137 | IndexError: pop from an empty deque with DeepSpeed ZeRO3 | ### System Info
- `transformers` version: 5.5.0.dev0
- Platform: Linux-5.15.0-1048-aws-x86_64-with-glibc2.31
- Python version: 3.10.18
- Huggingface_hub version: 1.8.0
- Safetensors version: 0.6.2
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: 0.18.8
- PyTorch version (accelerator?):... | closed | completed | false | 0 | [
"bug"
] | [] | 2026-03-31T07:38:10Z | 2026-04-13T16:38:52Z | 2026-04-13T16:38:52Z | MEMBER | null | 20260414T122001Z | 2026-04-14T12:20:01Z | albertvillanova | 8,515,462 | MDQ6VXNlcjg1MTU0NjI= | User | false |
huggingface/transformers | 4,178,863,303 | I_kwDOCUB6oc75FGTH | 45,141 | https://github.com/huggingface/transformers/issues/45141 | https://api.github.com/repos/huggingface/transformers/issues/45141 | [refactor] gpt-oss `eager_attention_forward` for modularity (Ex: models with dual eager attn: sink/no sink) | ### Feature request
Hi,
I'm finalizing a PR for integrating MiMo-V2, but for my `MiMoV2FlashAttention` class I wanted to reuse, for modularity, both the classic (no sink) `eager_attention_forward` from LLama and the perma sink one from Arthur in gpt-oss.
So that the full attention layers can benefit from SDPA/FLA2/f... | closed | completed | false | 1 | [
"Feature request"
] | [] | 2026-03-31T12:42:42Z | 2026-04-02T13:27:33Z | 2026-04-02T13:27:33Z | CONTRIBUTOR | null | 20260407T090028Z | 2026-04-07T09:00:28Z | casinca | 47,400,729 | MDQ6VXNlcjQ3NDAwNzI5 | User | false |
huggingface/transformers | 4,179,801,642 | I_kwDOCUB6oc75IrYq | 45,145 | https://github.com/huggingface/transformers/issues/45145 | https://api.github.com/repos/huggingface/transformers/issues/45145 | [Energy] N6 Arithmetic: 50-70% AI Training/Inference Energy Reduction — 17 Techniques with Code | ## Summary
**n=6 arithmetic reduces AI training and inference energy by 50-70%.** No hyperparameter search needed — all optimal values are mathematically predetermined from the unique solution to σ(n)·φ(n) = n·τ(n) ⟺ n = 6.
**Full Guide**: [AI Energy Savings Guide](https://github.com/need-singularity/n6-architecture/... | closed | completed | false | 0 | [] | [] | 2026-03-31T14:42:06Z | 2026-04-02T11:09:35Z | 2026-04-02T11:09:35Z | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | dancinlife | 44,921,882 | MDQ6VXNlcjQ0OTIxODgy | User | false |
huggingface/transformers | 4,179,820,107 | I_kwDOCUB6oc75Iv5L | 45,146 | https://github.com/huggingface/transformers/issues/45146 | https://api.github.com/repos/huggingface/transformers/issues/45146 | Allow for "pure" linear attention based Qwen3.5 models | ### Feature request
This feature requests proposes to allow for the creation of "pure" linear attention Qwen3.5 models. Which means that every layers should be allowed to be a Gated Deltanet token mixer.
Following code should therefore be "allowed":
```py
import torch
from transformers.models import Qwen3_5ForCaus... | closed | completed | false | 4 | [
"Feature request"
] | [] | 2026-03-31T14:44:05Z | 2026-04-02T11:17:06Z | 2026-04-02T11:17:06Z | CONTRIBUTOR | null | 20260407T090028Z | 2026-04-07T09:00:28Z | HallerPatrick | 22,773,355 | MDQ6VXNlcjIyNzczMzU1 | User | false |
huggingface/transformers | 4,181,308,817 | I_kwDOCUB6oc75ObWR | 45,151 | https://github.com/huggingface/transformers/issues/45151 | https://api.github.com/repos/huggingface/transformers/issues/45151 | [Energy] N6 Arithmetic: 50-70% AI Training/Inference Energy Reduction — 17 Techniques with Code | 🌍 **Open-source initiative to solve the global AI energy crisis.**
AI infrastructure energy consumption is doubling every year. This research provides mathematically proven techniques to cut training and inference energy by 50-70%, with no proprietary tools needed.
🔓 All code, proofs, and documentation are fully op... | closed | completed | false | 1 | [] | [] | 2026-03-31T18:10:50Z | 2026-04-02T11:13:13Z | 2026-04-02T11:09:09Z | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | dancinlife | 44,921,882 | MDQ6VXNlcjQ0OTIxODgy | User | false |
huggingface/transformers | 4,184,682,775 | I_kwDOCUB6oc75bTEX | 45,161 | https://github.com/huggingface/transformers/issues/45161 | https://api.github.com/repos/huggingface/transformers/issues/45161 | Only TP not working with GPT-OSS MoE model | ### System Info
python version: 3.10
transformers version: v5.1-release branch
torch version: 2.9.1+cu128
### Who can help?
While running my training script to finetune gpt-oss-20b model using Tensor Parallelism, the code is breaking at this line in [moe.py](https://github.com/huggingface/transformers/blob/3fa4da70f... | open | null | false | 2 | [
"bug"
] | [] | 2026-04-01T07:07:22Z | 2026-05-04T03:42:23Z | null | NONE | null | 20260504T060033Z | 2026-05-04T06:00:33Z | SharvariMedhe | 178,161,239 | U_kgDOCp6GVw | User | false |
huggingface/transformers | 4,184,987,992 | I_kwDOCUB6oc75cdlY | 45,162 | https://github.com/huggingface/transformers/issues/45162 | https://api.github.com/repos/huggingface/transformers/issues/45162 | Document limitations of `PreTrainedModel._can_set_*` source inspection logic | Hi! While reading the implementation of PreTrainedModel._can_set_*, I noticed that it relies on inspecting the module source file via __file__ and performing string-based checks.
This seems to work in typical cases, but there are a few scenarios where it may not behave as expected:
- The searched strings may appear i... | closed | completed | false | 3 | [] | [] | 2026-04-01T08:03:59Z | 2026-05-12T08:54:31Z | 2026-05-12T08:54:31Z | NONE | null | 20260512T120027Z | 2026-05-12T12:00:27Z | ElgoogUdiab | 28,426,666 | MDQ6VXNlcjI4NDI2NjY2 | User | false |
huggingface/transformers | 4,190,662,664 | I_kwDOCUB6oc75yHAI | 45,175 | https://github.com/huggingface/transformers/issues/45175 | https://api.github.com/repos/huggingface/transformers/issues/45175 | Feature request: Add EfficientViT-SAM (efficientvitsam) to Transformers | ### Feature request
[EfficientViT-SAM](https://github.com/mit-han-lab/efficientvit) combines MIT’s EfficientViT encoder with the SAM-style prompt encoder and mask decoder. It offers a lighter, faster alternative to ViT-based SAM for interactive segmentation while staying close to the same prompting and mask-decoding w... | open | null | false | 0 | [
"Feature request"
] | [] | 2026-04-02T00:31:07Z | 2026-04-02T00:31:07Z | null | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | masoudpz | 4,822,765 | MDQ6VXNlcjQ4MjI3NjU= | User | false |
huggingface/transformers | 4,191,057,386 | I_kwDOCUB6oc75znXq | 45,177 | https://github.com/huggingface/transformers/issues/45177 | https://api.github.com/repos/huggingface/transformers/issues/45177 | add DeepSeek-OCR2 | ### Model Description
[DeepSeek-OCR-2](https://huggingface.co/papers/2601.20552) is an OCR-specialized vision-language model proposed by the DeepSeek team.
The model uses a distinctive architecture:
- **Vision encoder**: SAM ViT-B
- **Hybrid attention encoder**: Qwen2-based, applying bidirectional attention over i... | closed | completed | false | 1 | [
"New model"
] | [] | 2026-04-02T02:43:49Z | 2026-04-30T14:13:41Z | 2026-04-30T14:13:41Z | CONTRIBUTOR | null | 20260501T113108Z | 2026-05-01T11:31:08Z | thisisiron | 23,303,033 | MDQ6VXNlcjIzMzAzMDMz | User | false |
huggingface/transformers | 4,192,752,849 | I_kwDOCUB6oc756FTR | 45,182 | https://github.com/huggingface/transformers/issues/45182 | https://api.github.com/repos/huggingface/transformers/issues/45182 | 🔒 Track: Pin GitHub Actions to commit SHAs | ## Tracking issue — Pin GitHub Actions to commit SHAs
This issue tracks the migration of all GitHub Actions workflow files to use pinned commit SHAs instead of mutable tags or branch names (e.g. \`v4\`, \`main\`).
**Why?**
Pinning to a SHA prevents supply chain attacks where a tag could be silently moved to point to ... | closed | completed | false | 0 | [
"actions-pin-sha"
] | [] | 2026-04-02T08:19:53Z | 2026-04-02T08:23:31Z | 2026-04-02T08:23:31Z | MEMBER | null | 20260407T090028Z | 2026-04-07T09:00:28Z | paulinebm | 155,966,238 | U_kgDOCUvbHg | User | false |
huggingface/transformers | 4,193,166,790 | I_kwDOCUB6oc757qXG | 45,183 | https://github.com/huggingface/transformers/issues/45183 | https://api.github.com/repos/huggingface/transformers/issues/45183 | [Bug] XOR logic for `input_ids`/`inputs_embeds` validation produces wrong or misleading error messages across multiple models | ### System Info
## Summary
Multiple models across the library use an XOR (`^`) condition to validate
`input_ids` and `inputs_embeds` inputs. This pattern has two distinct issues
depending on the model:
1. **Severe (11 models)**: XOR is paired with the error message
`"You cannot specify both ..."`, which is fac... | closed | completed | false | 0 | [
"bug",
"Code agent slop"
] | [] | 2026-04-02T09:17:32Z | 2026-04-02T11:56:39Z | 2026-04-02T11:56:33Z | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | XanxusCrypto | 30,135,657 | MDQ6VXNlcjMwMTM1NjU3 | User | false |
huggingface/transformers | 4,196,681,150 | I_kwDOCUB6oc76JEW- | 45,198 | https://github.com/huggingface/transformers/issues/45198 | https://api.github.com/repos/huggingface/transformers/issues/45198 | [BUG] Wav2Vec2 wav2vec2-lv-60-espeak-cv-ft: save_pretrained and tokenization fail | ### System Info
* `transformers` version: `5.5.0.dev0`
* Platform: `Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.39`
* Python version: `3.12.3`
* `huggingface_hub` version: `1.8.0`
* `safetensors` version: `0.7.0`
* `accelerate` version: `1.13.0`
* Accelerate config: `not found`
* DeepSpeed version: `no... | closed | completed | false | 0 | [
"bug"
] | [] | 2026-04-02T19:58:10Z | 2026-04-18T09:05:31Z | 2026-04-14T13:59:04Z | CONTRIBUTOR | null | 20260418T090534Z | 2026-04-18T09:05:34Z | harshaljanjani | 75,426,551 | MDQ6VXNlcjc1NDI2NTUx | User | false |
huggingface/transformers | 4,196,814,649 | I_kwDOCUB6oc76Jk85 | 45,200 | https://github.com/huggingface/transformers/issues/45200 | https://api.github.com/repos/huggingface/transformers/issues/45200 | [Gemma 4] mm_token_type_ids required for text-only fine-tuning - should default to zeros | ### System Info
transformers: 5.5.0.dev0 (installed from source)
torch: 2.8.0+cu128
trl: 1.0.0
peft: 0.18.2.dev0
Python: 3.12
OS: Linux (RunPod, Ubuntu 24.04)
GPU: NVIDIA B200 (192GB)
### Who can help?
@zucchini-nlp @ArthurZucker
### Information
- [ ] The official example scripts
- [x] My own modified scripts
... | closed | completed | false | 5 | [
"bug"
] | [] | 2026-04-02T20:26:24Z | 2026-04-22T10:44:25Z | 2026-04-22T10:44:25Z | NONE | null | 20260422T120052Z | 2026-04-22T12:00:52Z | dentity007 | 184,003,273 | U_kgDOCveqyQ | User | false |
huggingface/transformers | 4,197,078,693 | I_kwDOCUB6oc76Klal | 45,201 | https://github.com/huggingface/transformers/issues/45201 | https://api.github.com/repos/huggingface/transformers/issues/45201 | [Gemma 4] Support per-layer FlashAttention: FA2 for sliding layers, SDPA for global layers | ## Problem
Gemma 4 (26B-A4B) has a **hybrid attention architecture** where different layers use different head dimensions:
- **26 out of 30 layers** use **sliding window attention** with `head_dim=256` — fully compatible with FlashAttention 2
- **4 out of 30 layers** use **global attention** with `global_head_dim=512... | closed | completed | false | 4 | [] | [] | 2026-04-02T21:30:14Z | 2026-05-12T08:54:28Z | 2026-05-12T08:54:28Z | NONE | null | 20260512T120027Z | 2026-05-12T12:00:27Z | samuelazran | 2,499,928 | MDQ6VXNlcjI0OTk5Mjg= | User | false |
huggingface/transformers | 4,197,807,024 | I_kwDOCUB6oc76NXOw | 45,203 | https://github.com/huggingface/transformers/issues/45203 | https://api.github.com/repos/huggingface/transformers/issues/45203 | Add PolarQuant quantization: Hadamard-rotated Lloyd-Max optimal weights + KV cache | ## 🚀 Feature request
### Motivation
PolarQuant is a quantization method that uses **Walsh-Hadamard rotation + Lloyd-Max optimal centroids** for both weight compression and KV cache compression. It achieves better PPL per bit than existing methods because:
1. **Hadamard rotation** decorrelates weight/activation valu... | open | null | false | 19 | [] | [] | 2026-04-03T01:52:14Z | 2026-05-05T23:36:12Z | null | NONE | null | 20260506T000041Z | 2026-05-06T00:00:41Z | caiovicentino | 193,428,813 | U_kgDOC4d9TQ | User | false |
huggingface/transformers | 4,198,343,838 | I_kwDOCUB6oc76PaSe | 45,205 | https://github.com/huggingface/transformers/issues/45205 | https://api.github.com/repos/huggingface/transformers/issues/45205 | Gemma4: chat_template missing from tokenizer_config.json, requires manual loading from separate file | ## Description
The Gemma4 models (e.g. `google/gemma-4-E2B-it`) don't include `chat_template` in `tokenizer_config.json`. The chat template is shipped as a separate `chat_template.jinja` file instead.
This means the standard `tokenizer.apply_chat_template()` workflow fails out of the box:
```python
from transformers... | open | null | false | 7 | [] | [] | 2026-04-03T04:55:51Z | 2026-05-07T15:42:25Z | null | CONTRIBUTOR | null | 20260507T180032Z | 2026-05-07T18:00:32Z | w4nderlust | 349,256 | MDQ6VXNlcjM0OTI1Ng== | User | false |
huggingface/transformers | 4,198,345,190 | I_kwDOCUB6oc76Panm | 45,206 | https://github.com/huggingface/transformers/issues/45206 | https://api.github.com/repos/huggingface/transformers/issues/45206 | Gemma4: PLE (Per-Layer Embeddings) implementation is underdocumented and config is misleading | ## Description
I was implementing Gemma4 inference from scratch (in Rust) and the Per-Layer Embeddings (PLE) system was by far the hardest part to get right. The config fields are misleading, the embedding type is non-obvious, and the full pipeline involves several undocumented steps. Sharing this in case it helps oth... | closed | completed | false | 0 | [] | [] | 2026-04-03T04:56:17Z | 2026-04-14T17:01:13Z | 2026-04-14T17:01:13Z | CONTRIBUTOR | null | 20260414T200457Z | 2026-04-14T20:04:57Z | w4nderlust | 349,256 | MDQ6VXNlcjM0OTI1Ng== | User | false |
huggingface/transformers | 4,198,441,837 | I_kwDOCUB6oc76PyNt | 45,208 | https://github.com/huggingface/transformers/issues/45208 | https://api.github.com/repos/huggingface/transformers/issues/45208 | [Qwen3MoE] Potentially a bug on `Qwen3MoeSparseMoeBlock` | Hi,
I found a typing mismatch on [`Qwen3MoeSparseMoeBlock`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen3_moe/modeling_qwen3_moe.py#L275):
```python
class Qwen3MoeSparseMoeBlock(nn.Module):
def __init__(self, config: Qwen3MoeConfig):
super().__init__()
self.ex... | closed | completed | false | 2 | [] | [] | 2026-04-03T05:26:11Z | 2026-04-19T13:41:02Z | 2026-04-13T14:07:32Z | NONE | null | 20260419T140535Z | 2026-04-19T14:05:35Z | KbKuuhaku | 30,406,457 | MDQ6VXNlcjMwNDA2NDU3 | User | false |
huggingface/transformers | 4,199,586,186 | I_kwDOCUB6oc76UJmK | 45,216 | https://github.com/huggingface/transformers/issues/45216 | https://api.github.com/repos/huggingface/transformers/issues/45216 | [Regression] Qwen3.5 saved checkpoint is not correct with `save_pretrained` API since version 5.4.0 | ### System Info
transformers == 5.3.0 works well
transformers ==5.4.0 returns `Unexpected model.language_model.language_model.language_model.layers.7.self_attn.v_proj.weight in loaded safetensors file`
### Who can help?
@zucchini-nlp
### Information
- [ ] The official example scripts
- [x] My own modified scripts
... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-04-03T09:42:19Z | 2026-04-09T13:17:51Z | 2026-04-09T13:17:51Z | CONTRIBUTOR | null | 20260411T144729Z | 2026-04-11T14:47:29Z | xin3he | 83,260,933 | MDQ6VXNlcjgzMjYwOTMz | User | false |
huggingface/transformers | 4,202,205,162 | I_kwDOCUB6oc76eI_q | 45,229 | https://github.com/huggingface/transformers/issues/45229 | https://api.github.com/repos/huggingface/transformers/issues/45229 | Gemma4 31B-IT Multi-GPU inference CUDA OOM | ### System Info
### Description
I updated my transformers module to 5.5.0 from 4.53.0 to try `google/gemma-4-31B-it` model. I was using `meta-llama/Llama-3.3-70B-Instruct` for the same set of prompts. The Llama model is able to process the prompt without any problems despite occupying more VRAM than Gemma4. Gemma4 on ... | closed | completed | false | 21 | [
"bug"
] | [] | 2026-04-03T21:07:41Z | 2026-04-23T08:18:03Z | 2026-04-23T08:18:03Z | NONE | null | 20260423T120024Z | 2026-04-23T12:00:24Z | vaibhavBh-0 | 94,278,954 | U_kgDOBZ6VKg | User | false |
huggingface/transformers | 4,202,212,283 | I_kwDOCUB6oc76eKu7 | 45,230 | https://github.com/huggingface/transformers/issues/45230 | https://api.github.com/repos/huggingface/transformers/issues/45230 | Bug report | ### System Info
Ch
### Who can help?
_No response_
### Information
- [x] The official example scripts
- [x] My own modified scripts
### Tasks
- [x] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [x] #45231
### Reproduction
Vvvbbbbbx
### Expected behavior
<img width="495" he... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-04-03T21:10:02Z | 2026-04-03T21:13:54Z | 2026-04-03T21:13:35Z | NONE | null | 20260407T090028Z | 2026-04-07T09:00:28Z | kerrrang9214-tech | 242,015,866 | U_kgDODmzeeg | User | false |
huggingface/transformers | 4,202,223,249 | I_kwDOCUB6oc76eNaR | 45,231 | https://github.com/huggingface/transformers/issues/45231 | https://api.github.com/repos/huggingface/transformers/issues/45231 | My own task or dataset (give details below) | null | closed | completed | false | 0 | [] | [] | 2026-04-03T21:13:54Z | 2026-04-08T13:07:14Z | 2026-04-08T13:07:14Z | NONE | null | 20260411T144729Z | 2026-04-11T14:47:29Z | kerrrang9214-tech | 242,015,866 | U_kgDODmzeeg | User | false |
huggingface/transformers | 4,203,993,670 | I_kwDOCUB6oc76k9pG | 45,237 | https://github.com/huggingface/transformers/issues/45237 | https://api.github.com/repos/huggingface/transformers/issues/45237 | GPT-OSS-20B not work in AMD GPUs | ### System Info
GPT-OSS-20B does not work on Radeon GPUs.
I tested it in both the native environment and the Docker container rocm/pytorch:rocm7.2.1_ubuntu24.04_py3.12_pytorch_release_2.9.1.
I tried updating Triton, but it still didn't work.
I tried those versions of Triton, triton-rocm 3.6.0, 3.5.1+rocm (included in ... | open | null | false | 4 | [
"bug"
] | [] | 2026-04-04T07:20:50Z | 2026-05-01T00:51:47Z | null | CONTRIBUTOR | null | 20260501T113108Z | 2026-05-01T11:31:08Z | tanreinama | 51,933,889 | MDQ6VXNlcjUxOTMzODg5 | User | false |
huggingface/transformers | 4,204,346,376 | I_kwDOCUB6oc76mTwI | 45,239 | https://github.com/huggingface/transformers/issues/45239 | https://api.github.com/repos/huggingface/transformers/issues/45239 | 🚨 QA Observer Agent: Real-Time Architecture & Security Pattern Watcher (SCAFFOLD-WATCH) | ### Feature request
Proposing SCAFFOLD-WATCH — an observer agent to proactively surface architectural drift, security vulnerabilities (e.g. credential leaks, unparameterized SQL, agent drift) and redundant/repetitive developer work in real-time across PRs and developer sessions.
Systems like Transformers are highly ... | closed | completed | false | 1 | [] | [] | 2026-04-04T09:14:13Z | 2026-04-08T13:12:28Z | 2026-04-08T13:12:28Z | NONE | null | 20260411T144729Z | 2026-04-11T14:47:29Z | Insider77Circle | 160,362,522 | U_kgDOCY7wGg | User | false |
huggingface/transformers | 4,205,336,949 | I_kwDOCUB6oc76qFl1 | 45,242 | https://github.com/huggingface/transformers/issues/45242 | https://api.github.com/repos/huggingface/transformers/issues/45242 | [Gemma 4] `use_cache=False` corrupts attention computation, producing garbage logits | Gemma 4 has a bug where `use_cache=False` corrupts the attention computation, producing garbage logits. Every QLoRA tutorial sets `model.config.use_cache = False`, but this breaks Gemma 4 specifically.
When fine-tuning Gemma 4 (E2B-it in this situation) using standard QLoRA/LoRA workflows, the model produces garbage l... | closed | completed | false | 6 | [] | [] | 2026-04-04T16:48:25Z | 2026-04-10T12:07:11Z | 2026-04-09T08:17:52Z | NONE | null | 20260411T144729Z | 2026-04-11T14:47:29Z | siwoolol | 99,875,626 | U_kgDOBfP7Kg | User | false |
huggingface/transformers | 4,206,107,487 | I_kwDOCUB6oc76tBtf | 45,245 | https://github.com/huggingface/transformers/issues/45245 | https://api.github.com/repos/huggingface/transformers/issues/45245 | RuntimeError: number of categories cannot exceed 2^24 | ### System Info
### System Info
- `transformers` version: 5.2.0
- Platform: Linux (Ubuntu) / CUDA
### Information
- [x] The official example scripts
- [x] My own modified scripts
### Bug description
When using `model.generate()` with `do_sample=True` and a large `num_beams` (e.g., 128) on a model with a large vocab... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-04-05T00:42:05Z | 2026-05-08T11:25:13Z | 2026-05-08T11:25:13Z | NONE | null | 20260508T120022Z | 2026-05-08T12:00:22Z | Hzzone | 19,267,349 | MDQ6VXNlcjE5MjY3MzQ5 | User | false |
huggingface/transformers | 4,206,312,958 | I_kwDOCUB6oc76tz3- | 45,246 | https://github.com/huggingface/transformers/issues/45246 | https://api.github.com/repos/huggingface/transformers/issues/45246 | [Security/Feature] Deterministic substrate made modeling_utils.py stateful without modifying source — CJPI 100 · CVE-2025-32434 wrapped | Hi HuggingFace team,
I’m a solo developer. I built a deterministic, non-AI software evolution engine called CMPSBL® and I ran it on modeling_utils.py — your foundational training model utility layer.
I want to be direct about what happened:
The substrate wrapped the file in 217 seconds. It did not modify a single li... | closed | completed | false | 15 | [] | [] | 2026-04-05T03:06:11Z | 2026-04-11T16:40:17Z | 2026-04-08T13:14:48Z | NONE | null | 20260413T085906Z | 2026-04-13T08:59:06Z | SweetKenneth | 125,772,314 | U_kgDOB38iGg | User | false |
huggingface/transformers | 4,207,250,831 | I_kwDOCUB6oc76xY2P | 45,250 | https://github.com/huggingface/transformers/issues/45250 | https://api.github.com/repos/huggingface/transformers/issues/45250 | Flash Attention 2.0 | : NemotronHForCausalLM does not support Flash Attention 2.0 yet. Please request to add support where the model is hosted, on its model
hub page: https://huggingface.co//kaggle/input/models/metric/nemotron-3-nano-30b-a3b-bf16/transformers/default/1/discussions/new or in the
Transformers GitHub repo: ... | closed | completed | false | 1 | [] | [] | 2026-04-05T11:04:24Z | 2026-04-07T10:43:33Z | 2026-04-07T10:43:33Z | NONE | null | 20260411T144729Z | 2026-04-11T14:47:29Z | nicholasdudek | 196,255,629 | U_kgDOC7KfjQ | User | false |
huggingface/transformers | 4,209,709,972 | I_kwDOCUB6oc766xOU | 45,259 | https://github.com/huggingface/transformers/issues/45259 | https://api.github.com/repos/huggingface/transformers/issues/45259 | How can I use Gemma4's variable image resolution feature? | Hi,
I'd like to adjust the vision token budget. It appears that `default_output_length` in vision_config controls the length of the vision output tokens, but modifying it in `config.json` or loading the processor with `default_output_length=560` don't work.
Could you advise how to change the number of vision outputs? | closed | completed | false | 1 | [] | [] | 2026-04-06T05:14:28Z | 2026-04-14T10:52:32Z | 2026-04-14T10:52:32Z | NONE | null | 20260414T122001Z | 2026-04-14T12:20:01Z | ejlee95 | 36,064,879 | MDQ6VXNlcjM2MDY0ODc5 | User | false |
huggingface/transformers | 4,212,778,412 | I_kwDOCUB6oc77GeWs | 45,265 | https://github.com/huggingface/transformers/issues/45265 | https://api.github.com/repos/huggingface/transformers/issues/45265 | More permissive config parsing and validation | ### Feature request
Make more permissive `config.json`/`params.json` parsing / validation: cast int constants as float without warnings
### Motivation
E.g. when loading Leanstral (cf https://huggingface.co/mistralai/Leanstral-2603/discussions/7#69cfde05abe040f5323c6390):
```
Unrecognized keys in `rope_parameters` f... | closed | completed | false | 9 | [
"Feature request"
] | [] | 2026-04-06T16:28:50Z | 2026-04-13T13:14:41Z | 2026-04-13T13:14:41Z | NONE | null | 20260414T122001Z | 2026-04-14T12:20:01Z | vadimkantorov | 1,041,752 | MDQ6VXNlcjEwNDE3NTI= | User | false |
huggingface/transformers | 4,215,133,968 | I_kwDOCUB6oc77PdcQ | 45,276 | https://github.com/huggingface/transformers/issues/45276 | https://api.github.com/repos/huggingface/transformers/issues/45276 | [gemma4] resize_token_embeddings does not effect to embed_tokens_per_layer or output_embeddings | ### System Info
- `transformers` version: 5.5.0
- Platform: Linux-6.6.113+-x86_64-with-glibc2.35
- Python version: 3.12.13
- Huggingface_hub version: 1.8.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.... | closed | completed | false | 8 | [
"bug"
] | [] | 2026-04-07T02:55:03Z | 2026-04-15T11:15:23Z | 2026-04-15T11:15:23Z | CONTRIBUTOR | null | 20260415T224019Z | 2026-04-15T22:40:19Z | KoichiYasuoka | 15,098,598 | MDQ6VXNlcjE1MDk4NTk4 | User | false |
huggingface/transformers | 4,215,644,353 | I_kwDOCUB6oc77RaDB | 45,278 | https://github.com/huggingface/transformers/issues/45278 | https://api.github.com/repos/huggingface/transformers/issues/45278 | Many import errors after update from 4.57.0 to 5.5.0 | ### System Info
After updating Transformers from version 4.57.0 to version 5.5.0, I get the following import errors:
ImportError: cannot import name 'HybridCache' from 'transformers'
ImportError: cannot import name 'AutoModelForVision2Seq' from 'transformers'
ImportError: cannot import name 'PretrainedConfig' from 't... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-04-07T05:28:47Z | 2026-05-15T08:58:19Z | 2026-05-15T08:58:19Z | NONE | null | 20260515T120027Z | 2026-05-15T12:00:27Z | marchcat69 | 67,062,782 | MDQ6VXNlcjY3MDYyNzgy | User | false |
huggingface/transformers | 4,217,995,415 | I_kwDOCUB6oc77aYCX | 45,290 | https://github.com/huggingface/transformers/issues/45290 | https://api.github.com/repos/huggingface/transformers/issues/45290 | `apply_chat_template(tokenize=True)` crashes on assistant messages with tool calls and no content | ### System Info
Transformers 5.5..0
### Who can help?
@zucchini-nlp @Rocketknight1
### Information
- [ ] The official example scripts
- [x] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [x] My own task or dataset (give details below)
##... | closed | completed | false | 2 | [
"bug"
] | [] | 2026-04-07T13:29:49Z | 2026-04-13T19:02:32Z | 2026-04-13T19:02:32Z | MEMBER | null | 20260414T122001Z | 2026-04-14T12:20:01Z | qgallouedec | 45,557,362 | MDQ6VXNlcjQ1NTU3MzYy | User | false |
huggingface/transformers | 4,218,427,722 | I_kwDOCUB6oc77cBlK | 45,292 | https://github.com/huggingface/transformers/issues/45292 | https://api.github.com/repos/huggingface/transformers/issues/45292 | resize_token_embeddings does not effect to output_embeddings | ### System Info
- `transformers` version: 5.5.0
- Platform: Linux-6.6.113+-x86_64-with-glibc2.35
- Python version: 3.12.13
- Huggingface_hub version: 1.8.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.10... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-04-07T14:36:45Z | 2026-04-25T09:16:04Z | 2026-04-25T09:16:04Z | CONTRIBUTOR | null | 20260425T120019Z | 2026-04-25T12:00:19Z | KoichiYasuoka | 15,098,598 | MDQ6VXNlcjE1MDk4NTk4 | User | false |
huggingface/transformers | 4,219,591,188 | I_kwDOCUB6oc77gdoU | 45,295 | https://github.com/huggingface/transformers/issues/45295 | https://api.github.com/repos/huggingface/transformers/issues/45295 | Support Sequence Classification for Gemma 4 Models | ### Feature request
Add Gemma4ForSequenceClassification
### Motivation
Without this class, fine-tuning Gemma 4 on classification tasks requires manually adding a classification head, losing compatibility with AutoModelForSequenceClassification, Trainer, and the standard pipeline workflow.
### Your contribution
#45... | closed | completed | false | 2 | [
"Feature request"
] | [] | 2026-04-07T17:52:26Z | 2026-04-15T10:43:34Z | 2026-04-15T10:43:34Z | NONE | null | 20260415T224019Z | 2026-04-15T22:40:19Z | jesperschlegel | 17,431,950 | MDQ6VXNlcjE3NDMxOTUw | User | false |
huggingface/transformers | 4,221,012,848 | I_kwDOCUB6oc77l4tw | 45,304 | https://github.com/huggingface/transformers/issues/45304 | https://api.github.com/repos/huggingface/transformers/issues/45304 | NexusQuant: training-free KV cache compression (10-33x) via DynamicCache hooks | Hi,
Sharing a training-free KV cache compression approach we've been developing that hooks into DynamicCache. Might be useful for folks running into memory limits with long contexts.
**NexusQuant** compresses the KV cache by 10-33x by combining attention-based token eviction with E8 lattice vector quantization. It mo... | open | null | false | 2 | [] | [] | 2026-04-07T22:50:31Z | 2026-05-08T08:27:55Z | null | NONE | null | 20260508T120022Z | 2026-05-08T12:00:22Z | jagmarques | 32,335,502 | MDQ6VXNlcjMyMzM1NTAy | User | false |
huggingface/transformers | 4,221,154,706 | I_kwDOCUB6oc77mbWS | 45,305 | https://github.com/huggingface/transformers/issues/45305 | https://api.github.com/repos/huggingface/transformers/issues/45305 | Gradients not averaged by GAS when using DeepSpeed + model_accepts_loss_kwargs=True (Qwen3, Llama3, etc.) | ### System Info
- `transformers` version: 5.3.0
- Platform: Linux-5.14.0-427.33.1.el9_4.x86_64-x86_64-with-glibc2.34
- Python version: 3.11.15
- Huggingface_hub version: 1.6.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: - compute_environment: LOCAL_MACHINE
- distributed_ty... | closed | completed | false | 6 | [
"bug"
] | [] | 2026-04-07T23:31:43Z | 2026-04-13T14:32:25Z | 2026-04-13T14:32:25Z | CONTRIBUTOR | null | 20260414T122001Z | 2026-04-14T12:20:01Z | florian6973 | 70,778,912 | MDQ6VXNlcjcwNzc4OTEy | User | false |
huggingface/transformers | 4,221,321,450 | I_kwDOCUB6oc77nEDq | 45,306 | https://github.com/huggingface/transformers/issues/45306 | https://api.github.com/repos/huggingface/transformers/issues/45306 | Remove import * usage | Sorry but I want to revisit a closed issue #41669 here. That issue was automatically closed at the time without further responses. I hope we can reconsider the validity of that issue.
As I mentioned in that issue, I believe this is a serious enough problem. Actually, we're on the same page: we all want IDE autocomplet... | open | null | false | 2 | [] | [] | 2026-04-08T00:22:53Z | 2026-05-08T08:27:53Z | null | NONE | null | 20260508T120022Z | 2026-05-08T12:00:22Z | yueyinqiu | 18,749,772 | MDQ6VXNlcjE4NzQ5Nzcy | User | false |
huggingface/transformers | 4,222,028,568 | I_kwDOCUB6oc77pwsY | 45,307 | https://github.com/huggingface/transformers/issues/45307 | https://api.github.com/repos/huggingface/transformers/issues/45307 | Bug, Generation | ## Title
`AssistantToTargetTranslator` crashes with `AttributeError: 'map_input_embeddings'` when using assisted generation with cross-vocab models
## Description
When using assisted generation (`model.generate(assistant_model=...)`) with models that have different vocabulary sizes but share the same tokenizer famil... | closed | completed | false | 3 | [] | [] | 2026-04-08T04:07:37Z | 2026-04-10T09:30:25Z | 2026-04-10T09:30:25Z | CONTRIBUTOR | null | 20260411T144729Z | 2026-04-11T14:47:29Z | Regata3010 | 105,749,532 | U_kgDOBk2cHA | User | false |
huggingface/transformers | 4,222,532,106 | I_kwDOCUB6oc77rroK | 45,308 | https://github.com/huggingface/transformers/issues/45308 | https://api.github.com/repos/huggingface/transformers/issues/45308 | Feature request: Support evaluation every N epochs in TrainingArguments | ### Feature request
Currently, Trainer supports evaluation strategies:
- "epoch": evaluate every epoch
- "steps": evaluate every N steps
However, there is no built-in way to evaluate every N epochs (e.g., every 5 epochs).
This is particularly useful when:
- evaluation is computationally expensive
- running large-sca... | open | null | false | 3 | [
"Feature request"
] | [] | 2026-04-08T06:22:20Z | 2026-05-01T16:29:18Z | null | NONE | null | 20260501T180051Z | 2026-05-01T18:00:51Z | varuna-km-18267 | 201,051,200 | U_kgDOC_vMQA | User | false |
huggingface/transformers | 4,223,503,781 | I_kwDOCUB6oc77vY2l | 45,310 | https://github.com/huggingface/transformers/issues/45310 | https://api.github.com/repos/huggingface/transformers/issues/45310 | [BUG] transformers>=5.4.0, Qwen3.5 Moe from_pretrained error | ### System Info
```
import os
os.environ['CUDA_VISIBLE_DEVICS'] = '0'
from transformers import Qwen3_5ForConditionalGeneration, AutoTokenizer
model = Qwen3_5ForConditionalGeneration.from_pretrained('Qwen/Qwen3.5-35B-A3B')
model.save_pretrained('/root/Qwen3.5-35B-A3B', max_shard_size='10GB')
model = Qwen3_5ForCond... | closed | completed | false | 2 | [
"bug"
] | [] | 2026-04-08T09:29:48Z | 2026-04-09T13:26:04Z | 2026-04-09T13:26:04Z | CONTRIBUTOR | null | 20260411T144729Z | 2026-04-11T14:47:29Z | Jintao-Huang | 45,290,347 | MDQ6VXNlcjQ1MjkwMzQ3 | User | false |
huggingface/transformers | 4,224,244,253 | I_kwDOCUB6oc77yNod | 45,313 | https://github.com/huggingface/transformers/issues/45313 | https://api.github.com/repos/huggingface/transformers/issues/45313 | Qwen3.5: DeepSpeed ZeRO-3 fails to load weights for language_model | ## System Info
- `transformers` version: 5.4.0
- Platform: Linux (H200 x4)
- Python version: 3.12.0
- DeepSpeed version: 0.18.5
- PyTorch version: 2.8.0+cu128 (CUDA)
## Problem
When loading Qwen/Qwen3.5-27B (also tested with 9B) with DeepSpeed ZeRO-3, `language_model` parameters are reported as MISSING in the load rep... | open | null | false | 9 | [
"Should Fix"
] | [] | 2026-04-08T11:44:10Z | 2026-05-13T02:44:22Z | null | NONE | null | 20260513T060025Z | 2026-05-13T06:00:25Z | debOliveira | 48,807,586 | MDQ6VXNlcjQ4ODA3NTg2 | User | false |
huggingface/transformers | 4,225,839,225 | I_kwDOCUB6oc774TB5 | 45,322 | https://github.com/huggingface/transformers/issues/45322 | https://api.github.com/repos/huggingface/transformers/issues/45322 | [Model Request] Add EUPE (Efficient Universal Perception Encoder) by Meta AI | ### Model description
I would like to request adding support for EUPE (Efficient Universal Perception Encoder), a recent vision backbone released by Meta AI.
Paper: https://arxiv.org/abs/2603.22387
Code: https://github.com/facebookresearch/eupe
EUPE is a multi-purpose vision encoder trained via distillation from mul... | open | null | false | 2 | [
"New model"
] | [] | 2026-04-08T15:52:40Z | 2026-04-28T02:29:11Z | null | NONE | null | 20260428T060015Z | 2026-04-28T06:00:15Z | forensicmike | 44,839,768 | MDQ6VXNlcjQ0ODM5NzY4 | User | false |
huggingface/transformers | 4,226,684,674 | I_kwDOCUB6oc777hcC | 45,325 | https://github.com/huggingface/transformers/issues/45325 | https://api.github.com/repos/huggingface/transformers/issues/45325 | Qwen2.5-VL get_rope_index scales still-image temporal position_ids by tokens_per_second in transformers 5.3.0 | ### System Info
- `transformers` version: 5.3.0
- Platform: Linux-5.15.0-131-generic-x86_64-with-glibc2.39
- Python version: 3.12.3
- Huggingface_hub version: 1.9.0
- Safetensors version: 0.6.2
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerato... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-04-08T18:20:48Z | 2026-04-10T09:17:44Z | 2026-04-10T09:17:44Z | NONE | null | 20260411T144729Z | 2026-04-11T14:47:29Z | ayaan-fw | 197,808,338 | U_kgDOC8pQ0g | User | false |
huggingface/transformers | 4,229,234,189 | I_kwDOCUB6oc78FP4N | 45,331 | https://github.com/huggingface/transformers/issues/45331 | https://api.github.com/repos/huggingface/transformers/issues/45331 | [Gemma4] Bug: audio token missing newline separators in chat_template.jinja causes multimodal failure when image precedes audio | ## Bug Description
In `chat_template.jinja` for Gemma4, the image token has `\n\n` separators
but the audio token does not:
```jinja
{%- elif item['type'] == 'image' -%}
{{- '\n\n<|image|>\n\n' -}} ← has \n\n
{%- elif item['type'] == 'audio' -%}
{{- '<|audio|>' -}} ← missing \n\n
This causes the ... | closed | completed | false | 2 | [] | [] | 2026-04-09T03:35:46Z | 2026-05-11T14:32:04Z | 2026-05-11T14:32:04Z | NONE | null | 20260511T180035Z | 2026-05-11T18:00:35Z | LfWhat | 103,490,056 | U_kgDOBisiCA | User | false |
huggingface/transformers | 4,230,606,726 | I_kwDOCUB6oc78Ke-G | 45,335 | https://github.com/huggingface/transformers/issues/45335 | https://api.github.com/repos/huggingface/transformers/issues/45335 | [t5gemma] resize_token_embeddings does not effect to decoder.embed_tokens | ### System Info
- `transformers` version: 5.5.1
- Platform: Linux-6.6.113+-x86_64-with-glibc2.35
- Python version: 3.12.13
- Huggingface_hub version: 1.9.2
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.10... | closed | completed | false | 4 | [
"bug"
] | [] | 2026-04-09T08:11:47Z | 2026-04-16T16:02:08Z | 2026-04-16T16:02:08Z | CONTRIBUTOR | null | 20260416T222012Z | 2026-04-16T22:20:12Z | KoichiYasuoka | 15,098,598 | MDQ6VXNlcjE1MDk4NTk4 | User | false |
huggingface/transformers | 4,232,449,575 | I_kwDOCUB6oc78Rg4n | 45,341 | https://github.com/huggingface/transformers/issues/45341 | https://api.github.com/repos/huggingface/transformers/issues/45341 | A little bug in testing_utils.py | ### System Info
.
### Who can help?
I was trying to run integration tests on cpu for a new model addition PR then got hit by an error due to line 3220
https://github.com/huggingface/transformers/blob/2fae57f5da9b0108d0c4da2692f5a702e6fb8c02/src/transformers/testing_utils.py#L3215-L3235
The reason is that I have CU... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-04-09T13:05:35Z | 2026-05-11T03:13:27Z | 2026-05-11T03:13:27Z | CONTRIBUTOR | null | 20260511T060028Z | 2026-05-11T06:00:28Z | MHRDYN7 | 113,298,714 | U_kgDOBsDNGg | User | false |
huggingface/transformers | 4,238,183,287 | I_kwDOCUB6oc78nYt3 | 45,356 | https://github.com/huggingface/transformers/issues/45356 | https://api.github.com/repos/huggingface/transformers/issues/45356 | Regression in Kimi-K2.5 tokenizer from 5.3.0 to 5.4.0: incorrect codec handling and misleading fix_mistral_regex warning | ### System Info
- OS: Linux
- Python: 3.10.12
- Model/tokenizer: `moonshotai/Kimi-K2.5`
- `trust_remote_code=True`
### Who can help?
@ArthurZucker and @itazap
### Information
- [ ] The official example scripts
- [x] My own modified scripts
### Tasks
- [x] An officially supported task in the `examples` folder (s... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-04-10T09:42:26Z | 2026-04-13T16:43:05Z | 2026-04-13T15:16:25Z | NONE | null | 20260414T122001Z | 2026-04-14T12:20:01Z | Lander-Hatsune | 36,358,465 | MDQ6VXNlcjM2MzU4NDY1 | User | false |
huggingface/transformers | 4,238,326,785 | I_kwDOCUB6oc78n7wB | 45,357 | https://github.com/huggingface/transformers/issues/45357 | https://api.github.com/repos/huggingface/transformers/issues/45357 | [Regression] Qwen3.5 `save_pretrained` still saves incorrect visual encoder keys in 5.5.3 | ### System Info
- `transformers` version: 5.5.0, 5.5.3
- Platform: Linux (NVIDIA A100 80GB × 8)
- Python version: 3.12
- PyTorch version: 2.9.1+cu128
- CUDA version: 12.8
### Who can help?
_No response_
### Information
- [ ] The official example scripts
- [ ] My own modified scripts
### Tasks
- [ ] An officially... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-04-10T10:01:58Z | 2026-05-12T07:45:12Z | 2026-04-10T15:41:46Z | NONE | null | 20260512T120027Z | 2026-05-12T12:00:27Z | johnking0099 | 1,121,447 | MDQ6VXNlcjExMjE0NDc= | User | false |
huggingface/transformers | 4,240,172,193 | I_kwDOCUB6oc78u-Sh | 45,362 | https://github.com/huggingface/transformers/issues/45362 | https://api.github.com/repos/huggingface/transformers/issues/45362 | Qwen3.5-35B crashes with transformers chat | ### System Info
Using "transformers chat" with Qwen3.5-35B the moment a prompt is sent the server errors with an AttributeError
```
[snip]
File "/mnt/venvs/rocm6.4/lib64/python3.12/site-packages/transformers/cli/serving/chat_completion.py", line 174, in _streaming
queue, streamer = gen_manager.generate_streaming(... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-04-10T15:39:40Z | 2026-04-13T15:01:54Z | 2026-04-13T15:01:54Z | NONE | null | 20260414T122001Z | 2026-04-14T12:20:01Z | Sector14 | 7,682,567 | MDQ6VXNlcjc2ODI1Njc= | User | false |
huggingface/transformers | 4,243,389,991 | I_kwDOCUB6oc787P4n | 45,372 | https://github.com/huggingface/transformers/issues/45372 | https://api.github.com/repos/huggingface/transformers/issues/45372 | ImportError: cannot import name 'ReasoningEffort' from mistral_common breaks Gemma 4 processor loading in transformers 5.5.x / 5.6.0.dev0 | ### System Info
Environment
- transformers: 5.5.1 and 5.6.0.dev0 (git main as of ~2026-04-08)
- mistral-common: 1.9.1
- Python: 3.11.15
- Platform: macOS aarch64 (Apple Silicon)
- mlx-vlm: 0.4.x
Bug description
When loading any model via AutoProcessor (tested with google/gemma-4-31b-it and mlx-community/Qwen3.5-2B-4b... | open | null | false | 4 | [
"bug"
] | [] | 2026-04-11T06:34:13Z | 2026-05-15T10:31:54Z | null | NONE | null | 20260515T120027Z | 2026-05-15T12:00:27Z | Glademist | 40,288,837 | MDQ6VXNlcjQwMjg4ODM3 | User | false |
huggingface/transformers | 4,243,403,001 | I_kwDOCUB6oc787TD5 | 45,373 | https://github.com/huggingface/transformers/issues/45373 | https://api.github.com/repos/huggingface/transformers/issues/45373 | Add Gemma4ForSequenceClassification (missing from gemma4 module — Gemma 2/3 have it) | ### Feature request
<p style="white-space: pre-wrap; margin-top: 0.1em; margin-bottom: 0.2em; color: rgb(204, 204, 204); font-family: -apple-system, "system-ui", "Segoe UI", Roboto, sans-serif; font-size: 13px; font-style: normal; font-variant-ligatures: normal; font-variant-caps: normal; font-weig... | open | null | false | 6 | [
"Feature request"
] | [] | 2026-04-11T06:41:23Z | 2026-05-06T14:29:17Z | null | NONE | null | 20260506T180045Z | 2026-05-06T18:00:45Z | LarsKlawitter | 46,841,030 | MDQ6VXNlcjQ2ODQxMDMw | User | false |
huggingface/transformers | 4,244,561,953 | I_kwDOCUB6oc78_uAh | 45,375 | https://github.com/huggingface/transformers/issues/45375 | https://api.github.com/repos/huggingface/transformers/issues/45375 | Qwen3_5MoeVisionConfig missing deepstack_visual_indexes field — silently dropped by @strict | ### System Info
Transformers v5.5.0
The @strict decorator on Qwen3_5MoeVisionConfig (added in #41250) silently drops the deepstack_visual_indexes field during config loading, because it's not declared as a class attribute. However, every Qwen3.5 MoE model on HuggingFace ships
with this field in its config.json.
... | open | null | false | 2 | [
"bug"
] | [] | 2026-04-11T12:43:25Z | 2026-05-12T08:54:20Z | null | NONE | null | 20260512T120027Z | 2026-05-12T12:00:27Z | sharonyu-115 | 192,668,651 | U_kgDOC3vj6w | User | false |
huggingface/transformers | 4,244,594,758 | I_kwDOCUB6oc78_2BG | 45,376 | https://github.com/huggingface/transformers/issues/45376 | https://api.github.com/repos/huggingface/transformers/issues/45376 | [Bug] Loading gemma4 on transformers v4 raises misleading AttributeError: 'list' object has no attribute 'keys' instead of a clear version requirement error | ## System info
```
- transformers version: 4.x (any release before v5.5.0)
- Platform: macOS (Apple Silicon) / Linux
- Python version: 3.12 / 3.13
- Model: google/gemma-4-E4B-it (requires transformers >= 5.5.0)
```
## Background
`google/gemma-4-E4B-it` uses the `gemma4` architecture, which was introduced in transfor... | open | null | false | 5 | [
"bug"
] | [] | 2026-04-11T12:57:06Z | 2026-05-12T08:54:19Z | null | NONE | null | 20260512T120027Z | 2026-05-12T12:00:27Z | HARISH-CS-01 | 53,273,291 | MDQ6VXNlcjUzMjczMjkx | User | false |
huggingface/transformers | 4,245,939,857 | I_kwDOCUB6oc79E-aR | 45,381 | https://github.com/huggingface/transformers/issues/45381 | https://api.github.com/repos/huggingface/transformers/issues/45381 | transformers==5.3.0, qwen2.5-vl video input vision_position_ids seems to be wrong | ### System Info
transformers == 5.3.0
[But the bug seems to be with any transformers >= 5.3.0, even in the current main branch]
qwen_vl_utils == 0.0.14
Python 3.12.4
Cuda 12.6
### Who can help?
@zucchini-nlp
### Information
- [x] The official example scripts
- [ ] My own modified scripts
### Tasks
- [ ] An offi... | closed | completed | false | 7 | [
"bug"
] | [] | 2026-04-11T22:43:30Z | 2026-04-14T13:02:14Z | 2026-04-14T13:02:14Z | NONE | null | 20260414T151202Z | 2026-04-14T15:12:02Z | bicheng-xu | 20,759,548 | MDQ6VXNlcjIwNzU5NTQ4 | User | false |
huggingface/transformers | 4,251,348,974 | I_kwDOCUB6oc79Zm_u | 45,390 | https://github.com/huggingface/transformers/issues/45390 | https://api.github.com/repos/huggingface/transformers/issues/45390 | CLIPTextModel / CLIPVisionModel fail to load old checkpoints after architecture flattening | ## Description
After the recent refactoring that flattened `CLIPTextModel` (removed the `self.text_model` wrapper) and `CLIPVisionModel` (removed the `self.vision_model` wrapper), old checkpoints that were saved with the nested structure can no longer be loaded correctly.
All weights end up randomly initialized becau... | open | reopened | false | 2 | [] | [] | 2026-04-13T04:50:21Z | 2026-05-13T08:56:17Z | null | MEMBER | null | 20260513T120045Z | 2026-05-13T12:00:45Z | sayakpaul | 22,957,388 | MDQ6VXNlcjIyOTU3Mzg4 | User | false |
huggingface/transformers | 4,253,119,364 | I_kwDOCUB6oc79gXOE | 45,397 | https://github.com/huggingface/transformers/issues/45397 | https://api.github.com/repos/huggingface/transformers/issues/45397 | [BUG] gemma-4 zero3 from_pretrained | ### System Info
<img width="946" height="361" alt="Image" src="https://github.com/user-attachments/assets/193f1646-b90a-4d8b-a4ff-db8b252133ef" />
https://github.com/modelscope/ms-swift/issues/9078
google/gemma-4-E4B-it
zero2 works fine, zero3 does not.
zero2:
<img width="876" height="108" alt="Image" src="htt... | closed | completed | false | 4 | [
"bug"
] | [] | 2026-04-13T09:25:04Z | 2026-04-17T13:59:12Z | 2026-04-17T13:59:12Z | CONTRIBUTOR | null | 20260417T180542Z | 2026-04-17T18:05:42Z | Jintao-Huang | 45,290,347 | MDQ6VXNlcjQ1MjkwMzQ3 | User | false |
huggingface/transformers | 4,253,559,298 | I_kwDOCUB6oc79iCoC | 45,399 | https://github.com/huggingface/transformers/issues/45399 | https://api.github.com/repos/huggingface/transformers/issues/45399 | Fallback to kernels-community/flash-attn2 is blocked by other checks when fa2 is not installed | ### System Info
- `transformers` version: 5.5.3
- Platform: Linux-5.4.0-216-generic-x86_64-with-glibc2.31
- Python version: 3.10.0
- Huggingface_hub version: 1.10.1
- Safetensors version: 0.5.2
- Accelerate version: 1.6.0
- Accelerate config: - compute_environment: LOCAL_MACHINE
- distributed_type: MULTI_GP... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-04-13T10:20:19Z | 2026-04-14T02:16:52Z | 2026-04-14T02:16:52Z | CONTRIBUTOR | null | 20260414T122001Z | 2026-04-14T12:20:01Z | efsotr | 104,755,879 | U_kgDOBj5ypw | User | false |
huggingface/transformers | 4,255,127,163 | I_kwDOCUB6oc79oBZ7 | 45,405 | https://github.com/huggingface/transformers/issues/45405 | https://api.github.com/repos/huggingface/transformers/issues/45405 | `MIN_PEFT_VERSION` bumped to 0.18.2 which is not yet released on PyPI | ### System Info
- `transformers` version: 5.6.0.dev0 (post-commit c585eeaa65)
- Platform: Linux-5.14.0-570.12.1.el9_6.x86_64-x86_64-with-glibc2.34
- Python version: 3.11.15
- Huggingface_hub version: 1.9.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: n... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-04-13T13:31:01Z | 2026-04-13T13:59:24Z | 2026-04-13T13:59:24Z | NONE | null | 20260414T122001Z | 2026-04-14T12:20:01Z | artem-spector | 10,895,869 | MDQ6VXNlcjEwODk1ODY5 | User | false |
huggingface/transformers | 4,255,361,302 | I_kwDOCUB6oc79o6kW | 45,406 | https://github.com/huggingface/transformers/issues/45406 | https://api.github.com/repos/huggingface/transformers/issues/45406 | transformers serve crashes with AttributeError: 'Gemma4Processor' object has no attribute '_tokenizer' | ### System Info
transformers version: 5.5.3
Python version: 3.12.3
PyTorch version: 2.11.0+cu130
Platform: Linux (Ubuntu 24.04)
### Who can help?
@ArthurZucker ? idk
I will open an PR once i find time. I "fixed" it locally for now.
### Information
- [ ] The official example scripts
- [ ] My own modified scripts
#... | closed | completed | false | 4 | [
"bug"
] | [] | 2026-04-13T13:57:05Z | 2026-04-13T15:42:10Z | 2026-04-13T15:42:10Z | NONE | null | 20260414T122001Z | 2026-04-14T12:20:01Z | asdat3 | 29,544,382 | MDQ6VXNlcjI5NTQ0Mzgy | User | false |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.