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 | 4,560,498,294 | I_kwDOCUB6oc8AAAABD9Oudg | 46,315 | https://github.com/huggingface/transformers/issues/46315 | https://api.github.com/repos/huggingface/transformers/issues/46315 | Regression: `convert_tokens_to_ids` is much slower in v5 than v4 for slow tokenizers with many added tokens | ### System Info
- `transformers` version: 5.9.0
- Platform: Linux-6.8.0-117-generic-x86_64-with-glibc2.39
- Python version: 3.12.13
- PyTorch version (accelerator?): 2.8.0+cu128 (NA)
### Who can help?
@itazap
### Information
- [ ] The official example scripts
- [x] My own modified scripts
### Tasks
- [ ] An off... | closed | completed | false | 2 | [
"bug"
] | [] | 2026-06-01T05:04:20Z | 2026-06-03T16:24:41Z | 2026-06-03T16:24:41Z | NONE | null | 20260603T180028Z | 2026-06-03T18:00:28Z | ichizok | 943,423 | MDQ6VXNlcjk0MzQyMw== | User | false |
huggingface/transformers | 4,563,453,492 | I_kwDOCUB6oc8AAAABEADGNA | 46,326 | https://github.com/huggingface/transformers/issues/46326 | https://api.github.com/repos/huggingface/transformers/issues/46326 | Bug in continue_final_message for tokenizer based on mistral_common | ### System Info
- `transformers` version: 5.5.3
- Platform: Linux-5.14.0-427.81.1.el9_4.x86_64-x86_64-with-glibc2.34
- Python version: 3.11.14
- Huggingface_hub version: 1.11.0
- Safetensors version: 0.7.0
- Accelerate version: not installed
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.10.0+c... | closed | completed | false | 11 | [
"bug"
] | [] | 2026-06-01T13:25:46Z | 2026-06-03T15:13:20Z | 2026-06-03T15:13:20Z | NONE | null | 20260603T180028Z | 2026-06-03T18:00:28Z | ntnq4 | 83,763,813 | MDQ6VXNlcjgzNzYzODEz | User | false |
huggingface/transformers | 4,564,300,660 | I_kwDOCUB6oc8AAAABEA2zdA | 46,328 | https://github.com/huggingface/transformers/issues/46328 | https://api.github.com/repos/huggingface/transformers/issues/46328 | SequenceFeatureExtractor.pad wasting time converting numpy array to list of numpy arrays | ### System Info
This is environment agnositic. (also `transformers env` throws an error)
### 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 ta... | closed | completed | false | 0 | [
"bug"
] | [] | 2026-06-01T15:25:55Z | 2026-06-02T13:08:16Z | 2026-06-02T13:08:16Z | NONE | null | 20260602T180035Z | 2026-06-02T18:00:35Z | bolshoytoster | 91,278,344 | MDQ6VXNlcjkxMjc4MzQ0 | User | false |
huggingface/transformers | 4,575,318,786 | I_kwDOCUB6oc8AAAABELXTAg | 46,358 | https://github.com/huggingface/transformers/issues/46358 | https://api.github.com/repos/huggingface/transformers/issues/46358 | [ExecuTorch] convert_and_export_with_cache fails for vision-language and audio models | ## Description
`convert_and_export_with_cache` and `TorchExportableModuleWithStaticCache.forward`
currently only support text-only models. Attempting to export a vision-language model
(e.g. LLaVA, PaliGemma) or audio model (e.g. Whisper) fails because:
1. `TorchExportableModuleWithStaticCache.forward` does not acce... | open | null | false | 1 | [] | [] | 2026-06-02T22:09:56Z | 2026-07-03T09:00:10Z | null | NONE | null | 20260703T120023Z | 2026-07-03T12:00:23Z | ananyajoshi0203-cpu | 274,597,891 | U_kgDOEF4IAw | User | false |
huggingface/transformers | 4,577,103,126 | I_kwDOCUB6oc8AAAABENENFg | 46,365 | https://github.com/huggingface/transformers/issues/46365 | https://api.github.com/repos/huggingface/transformers/issues/46365 | I couldn't import Trainer from transformers until I uninstalled datasets. | ### System Info
transformers: 4.46.3
torch: 2.10.0+cu126
Python: 3.10.11 | packaged by Anaconda, Inc. | (main, May 16 2023, 00:55:32) [MSC v.1916 64 bit (AMD64)]
OS: Windows-10-10.0.26200-SP0
datasets: 4.8.5
VS Code: 1.118.1 (system setup)
### Who can help?
@SunMarc
### Information
- [ ] The official example scrip... | open | null | false | 6 | [
"bug"
] | [] | 2026-06-03T05:19:06Z | 2026-07-03T09:00:08Z | null | NONE | null | 20260703T120023Z | 2026-07-03T12:00:23Z | OneGneralf | 75,012,787 | MDQ6VXNlcjc1MDEyNzg3 | User | false |
huggingface/transformers | 4,577,783,811 | I_kwDOCUB6oc8AAAABENtwAw | 46,369 | https://github.com/huggingface/transformers/issues/46369 | https://api.github.com/repos/huggingface/transformers/issues/46369 | Memory overflow occurs when calling the sam3-litetext model | ### System Info
from io import BytesIO
import httpx
from PIL import Image
from transformers import AutoModel, AutoProcessor
model = AutoModel.from_pretrained("yonigozlan/sam3-litetext-s0", device_map="auto")
processor = AutoProcessor.from_pretrained("yonigozlan/sam3-litetext-s0")
image_url = "http://images.cocoda... | open | null | false | 3 | [
"bug"
] | [] | 2026-06-03T07:25:02Z | 2026-07-03T09:00:06Z | null | NONE | null | 20260703T120023Z | 2026-07-03T12:00:23Z | mcmingchang | 37,234,299 | MDQ6VXNlcjM3MjM0Mjk5 | User | false |
huggingface/transformers | 2,362,887,604 | I_kwDOCUB6oc6M1s20 | 31,501 | https://github.com/huggingface/transformers/issues/31501 | https://api.github.com/repos/huggingface/transformers/issues/31501 | FineWeb SLM Training doesn't start | ### System Info
This is my dev env:
https://github.com/abhinand5/runpod-utils/blob/main/docker/torch-lm-dev/Dockerfile
Using the latest docker.
Torch 2.3.1
CUDA 12.1
### Who can help?
@ArthurZucker @younesbelkada
### Information
- [X] The official example scripts
- [X] My own modified scripts
### Tasks
... | closed | completed | false | 7 | [
"trainer",
"Good Second Issue"
] | [] | 2024-06-19T17:50:19Z | 2026-06-03T11:23:31Z | 2026-06-03T11:23:30Z | NONE | null | 20260603T120101Z | 2026-06-03T12:01:01Z | abhinand5 | 35,622,449 | MDQ6VXNlcjM1NjIyNDQ5 | User | false |
huggingface/transformers | 4,577,847,348 | I_kwDOCUB6oc8AAAABENxoNA | 46,370 | https://github.com/huggingface/transformers/issues/46370 | https://api.github.com/repos/huggingface/transformers/issues/46370 | Deprecate CompileableContextVar in favor of plain ContextVar with upcoming Dynamo native support | ### Feature request
We are planning to propose native `contextvars.ContextVar.get()` and `.set()` support to TorchDynamo.
Once this lands in TorchDynamo, CompileableContextVar can be replaced with a plain ContextVar that works directly inside torch.compile - no wrapper class, no compile-time special-casing, and threa... | open | null | false | 6 | [
"Feature request"
] | [] | 2026-06-03T07:35:59Z | 2026-06-15T02:35:36Z | null | NONE | null | 20260615T060048Z | 2026-06-15T06:00:48Z | ayushsatyam146 | 55,888,723 | MDQ6VXNlcjU1ODg4NzIz | User | false |
huggingface/transformers | 4,579,798,913 | I_kwDOCUB6oc8AAAABEPovgQ | 46,382 | https://github.com/huggingface/transformers/issues/46382 | https://api.github.com/repos/huggingface/transformers/issues/46382 | Propagation of output_hidden_states to sub_configs | ### Feature request
Hi !
I was wondering if it was possible to automatically propagate the `output_hidden_states` attributes to `sub_configs`. For some pre-trained models such as the multimodal ones, this could be convenient to avoid manually mentioning the configs of each modality.
Thanks :)
### Motivation
I use ... | open | null | false | 5 | [
"Feature request"
] | [] | 2026-06-03T12:26:39Z | 2026-06-04T12:47:22Z | null | NONE | null | 20260604T180024Z | 2026-06-04T18:00:24Z | julien-gadonneix | 145,470,783 | U_kgDOCKu1Pw | User | false |
huggingface/transformers | 4,583,620,097 | I_kwDOCUB6oc8AAAABETR-AQ | 46,389 | https://github.com/huggingface/transformers/issues/46389 | https://api.github.com/repos/huggingface/transformers/issues/46389 | AutoProcessor import fails with NVIDIA torch 2.7 missing torch.float8_e8m0fnu | ### System Info
- transformers: 5.10.1
- torch: 2.7.0a0+7c8ec84dab.nv25.03
- Python: 3.12
- CUDA runner: CUDA 12.8.1.012 / NVIDIA driver 580.126.09
Transformers metadata still lists `torch>=2.4` for the torch/testing extras.
### Reproduction
In an environment with `transformers==5.10.1` and NVIDIA torch `2.7.0a0+7c... | closed | completed | false | 2 | [] | [] | 2026-06-03T21:03:36Z | 2026-06-05T08:42:23Z | 2026-06-05T08:42:23Z | NONE | null | 20260605T120123Z | 2026-06-05T12:01:23Z | doctorpangloss | 2,229,300 | MDQ6VXNlcjIyMjkzMDA= | User | false |
huggingface/transformers | 4,585,846,705 | I_kwDOCUB6oc8AAAABEVZ3sQ | 46,396 | https://github.com/huggingface/transformers/issues/46396 | https://api.github.com/repos/huggingface/transformers/issues/46396 | `added_tokens_encoder` rebuilds the entire added-token map on every access |
### Who can help?
@itazap
### Summary
Follow-up to #46323, which fixed the worst case: `_convert_token_to_id_with_added_voc` was reading the `added_tokens_encoder` property, which rebuilds and re-sorts the whole added-token map on every access. The property itself still rebuilds on every read, and a few call sites ... | closed | completed | false | 0 | [] | [] | 2026-06-04T04:40:19Z | 2026-06-22T11:01:46Z | 2026-06-22T11:01:46Z | CONTRIBUTOR | null | 20260622T120018Z | 2026-06-22T12:00:18Z | ishan-1010 | 98,383,932 | U_kgDOBd04PA | User | false |
huggingface/transformers | 4,586,458,991 | I_kwDOCUB6oc8AAAABEV_Pbw | 46,399 | https://github.com/huggingface/transformers/issues/46399 | https://api.github.com/repos/huggingface/transformers/issues/46399 | Qwen3.5: use_kernels=True crashes with ValueError on apply_rotary_pos_emb registration - missing decorator | ### System Info
- `transformers` version: `b665f3ab02a5b83eecc102079c4d86544e9d2d11`
- Platform: Linux
- Python version: 3.13
- `kernels` version: 0.15.1
- `torch` version: 2.11.0+cu130
- `torch.version.cuda`: 13.0
### Who can help?
@MekkCyber @drbh @ArthurZucker
### Information
- [ ] The official example scripts
... | closed | completed | false | 11 | [
"Good Difficult Issue",
"bug"
] | [] | 2026-06-04T06:39:10Z | 2026-06-05T14:39:00Z | 2026-06-05T14:39:00Z | NONE | null | 20260605T180026Z | 2026-06-05T18:00:26Z | rycerzes | 59,965,507 | MDQ6VXNlcjU5OTY1NTA3 | User | false |
huggingface/transformers | 4,586,957,923 | I_kwDOCUB6oc8AAAABEWdsYw | 46,402 | https://github.com/huggingface/transformers/issues/46402 | https://api.github.com/repos/huggingface/transformers/issues/46402 | Sam3Model text encoder weights are not loaded in transformers 5.10.1 | `Sam3Model.from_pretrained("facebook/sam3")` works with transformers 5.9, but in 5.10.1 the SAM3 text encoder weights are not loaded. They are reported as both unexpected and missing, which means the text encoder gets randomly initialized.
### Reproduction
```python
from transformers import Sam3Model, __version__
pr... | closed | completed | false | 2 | [
"bug"
] | [] | 2026-06-04T08:02:01Z | 2026-06-04T14:37:22Z | 2026-06-04T14:37:22Z | NONE | null | 20260604T180024Z | 2026-06-04T18:00:24Z | MalteEbner | 20,324,507 | MDQ6VXNlcjIwMzI0NTA3 | User | false |
huggingface/transformers | 4,589,915,878 | I_kwDOCUB6oc8AAAABEZSO5g | 46,421 | https://github.com/huggingface/transformers/issues/46421 | https://api.github.com/repos/huggingface/transformers/issues/46421 | # Cache lazy initialization leads to spurious recompilation with chunked prefill | ### System Info
When #39797 simplified the cache management, it introduced the cache lazy optimization, assuming that there would always
be an eager forward call to trigger the initialization (typically the prefill).
This is however not true with chunked prefill, and it triggers a recompile on a second call to genera... | closed | completed | false | 4 | [
"bug"
] | [
"dacorvo"
] | 2026-06-04T15:21:20Z | 2026-07-05T16:31:55Z | 2026-07-05T16:31:55Z | MEMBER | null | 20260705T180020Z | 2026-07-05T18:00:20Z | dacorvo | 1,910,518 | MDQ6VXNlcjE5MTA1MTg= | User | false |
huggingface/transformers | 2,036,541,854 | I_kwDOCUB6oc55Yyme | 27,957 | https://github.com/huggingface/transformers/issues/27957 | https://api.github.com/repos/huggingface/transformers/issues/27957 | XLMRoberta with Flash Attention 2 | ### System Info
- transformers version: 4.36.0
- Platform: Linux-4.19.0-22-amd64-x86_64-with-glibc2.31
- Python version: 3.10.13
- Huggingface_hub version: 0.19.4
- Safetensors version: 0.4.0
- Accelerate version: 0.24.1
- Accelerate config: not found
- PyTorch version (GPU?): 2.0.1+cu117 (True)
- Tensorflow... | closed | completed | false | 6 | [
"Good Second Issue",
"Feature request"
] | [] | 2023-12-11T21:16:49Z | 2026-06-05T11:18:37Z | 2026-06-05T11:18:37Z | NONE | null | 20260605T120123Z | 2026-06-05T12:01:23Z | IvanPy96 | 64,599,936 | MDQ6VXNlcjY0NTk5OTM2 | User | false |
huggingface/transformers | 4,590,792,633 | I_kwDOCUB6oc8AAAABEaHvuQ | 46,424 | https://github.com/huggingface/transformers/issues/46424 | https://api.github.com/repos/huggingface/transformers/issues/46424 | Cache config max_cache_length only honored for Quantized Cache | ### System Info
`GenerationConfig.cache_config` is documented as carrying "arguments used in the key-value cache class", but on the static-cache path it is silently ignored. A user setting `cache_config={"max_cache_len": N}` together with `cache_implementation="static"` gets a cache sized to `max_new_tokens + input_le... | closed | completed | false | 3 | [
"bug"
] | [
"dacorvo"
] | 2026-06-04T17:27:14Z | 2026-07-05T16:31:21Z | 2026-07-05T16:31:21Z | MEMBER | null | 20260705T180020Z | 2026-07-05T18:00:20Z | dacorvo | 1,910,518 | MDQ6VXNlcjE5MTA1MTg= | User | false |
huggingface/transformers | 4,592,517,477 | I_kwDOCUB6oc8AAAABEbxBZQ | 46,426 | https://github.com/huggingface/transformers/issues/46426 | https://api.github.com/repos/huggingface/transformers/issues/46426 | Add KORMo model | ### Model description
I'd like to add **KORMo** (Korean Open Reasoning Model) to Transformers as a native model.
KORMo is a fully open bilingual (Korean–English) LLM — weights, training code and training data
are all openly released. Paper: https://huggingface.co/papers/2510.09426
Architecturally, KORMo is identical... | open | null | false | 1 | [] | [] | 2026-06-04T22:13:07Z | 2026-07-05T08:58:37Z | null | NONE | null | 20260705T120019Z | 2026-07-05T12:00:19Z | mjkmain | 72,269,271 | MDQ6VXNlcjcyMjY5Mjcx | User | false |
huggingface/transformers | 4,594,611,161 | I_kwDOCUB6oc8AAAABEdwz2Q | 46,431 | https://github.com/huggingface/transformers/issues/46431 | https://api.github.com/repos/huggingface/transformers/issues/46431 | ValueError: Cannot use apply_chat_template because this processor does not have a chat template. | emmm, I get this error when using [gemma-4-12B](https://huggingface.co/google/gemma-4-12B)
transformers=5.10.1
ValueError: Cannot use apply_chat_template because this processor does not have a chat template.
look forward to your reply. Thanks very much.
The code is here
---
```
from transformers import AutoProcessor... | closed | completed | false | 4 | [] | [] | 2026-06-05T05:59:49Z | 2026-06-06T10:25:24Z | 2026-06-06T10:25:24Z | NONE | null | 20260606T120018Z | 2026-06-06T12:00:18Z | AlphaNext | 35,558,843 | MDQ6VXNlcjM1NTU4ODQz | User | false |
huggingface/transformers | 4,596,186,506 | I_kwDOCUB6oc8AAAABEfQ9ig | 46,441 | https://github.com/huggingface/transformers/issues/46441 | https://api.github.com/repos/huggingface/transformers/issues/46441 | Qwen3.5 crash with `cache_implementation="static"` — `KeyError: 'linear_attention'` | ### System Info
transformers **5.10.1** (latest release)
`generate()` pre-builds attention masks for compilable caches via `create_masks_for_generate`, which maps each `config.layer_types` entry through `LAYER_PATTERN_TO_MASK_FUNCTION_MAPPING`.
Hybrid linear-attention models (Qwen3.5, Qwen3-Next, MiniMax, …) have `l... | open | null | false | 1 | [] | [] | 2026-06-05T10:26:44Z | 2026-06-06T15:10:46Z | null | MEMBER | null | 20260606T180020Z | 2026-06-06T18:00:20Z | dacorvo | 1,910,518 | MDQ6VXNlcjE5MTA1MTg= | User | false |
huggingface/transformers | 4,596,149,400 | I_kwDOCUB6oc8AAAABEfOsmA | 46,439 | https://github.com/huggingface/transformers/issues/46439 | https://api.github.com/repos/huggingface/transformers/issues/46439 | Static cache `early_initialization` corrupts linear-attention layers (wrong shapes → `RuntimeError`) | ### System Info
transformers **5.10.1** (latest release)
`Cache.early_initialization(batch_size, num_heads, head_dim, dtype, device)` pre-allocates all cache layers ahead of time, assuming there is a helper registered to precompute the cache shape available for the layer type.
But there are no registered helpers for... | open | null | false | 0 | [] | [] | 2026-06-05T10:20:32Z | 2026-06-05T10:47:09Z | null | MEMBER | null | 20260605T120123Z | 2026-06-05T12:01:23Z | dacorvo | 1,910,518 | MDQ6VXNlcjE5MTA1MTg= | User | false |
huggingface/transformers | 4,596,690,333 | I_kwDOCUB6oc8AAAABEfvtnQ | 46,443 | https://github.com/huggingface/transformers/issues/46443 | https://api.github.com/repos/huggingface/transformers/issues/46443 | Vision RoPE refactor | ### Feature request
Vision RoPE layer/module should return the final positional embedding, following https://github.com/huggingface/transformers/pull/41992#discussion_r3361589165 @zucchini-nlp @ArthurZucker
### Motivation
We already create a class for it so might as well do the full work.
### Your contribution
Wi... | open | null | false | 7 | [
"Feature request"
] | [
"IlyasMoutawwakil"
] | 2026-06-05T11:52:58Z | 2026-06-11T09:53:14Z | null | MEMBER | null | 20260611T120117Z | 2026-06-11T12:01:17Z | IlyasMoutawwakil | 57,442,720 | MDQ6VXNlcjU3NDQyNzIw | User | false |
huggingface/transformers | 2,784,472,695 | I_kwDOCUB6oc6l9653 | 35,667 | https://github.com/huggingface/transformers/issues/35667 | https://api.github.com/repos/huggingface/transformers/issues/35667 | Improve Guidance for Using DDP in `examples/pytorch` | ### Feature request
The examples in `examples/pytorch/` (e.g., [semantic-segmentation](https://github.com/huggingface/transformers/tree/main/examples/pytorch/semantic-segmentation)) would benefit from clearer guidance on how to use Distributed Data Parallel (DDP) in trainer version.
### Motivation
I modified the tr... | closed | completed | false | 3 | [
"Feature request"
] | [] | 2025-01-13T16:08:22Z | 2026-06-08T11:41:28Z | 2026-06-08T11:41:28Z | NONE | null | 20260608T120020Z | 2026-06-08T12:00:20Z | caojiaolong | 18,239,971 | MDQ6VXNlcjE4MjM5OTcx | User | false |
huggingface/transformers | 4,601,803,187 | I_kwDOCUB6oc8AAAABEknxsw | 46,458 | https://github.com/huggingface/transformers/issues/46458 | https://api.github.com/repos/huggingface/transformers/issues/46458 | Suggestion: Give a warning for 4D attention mask which use float dtype but in 0.0 and 1.0 term (instead of -inf and zero). | ### Feature request
The title said it all.
### Motivation
Let's say I have a code like this: (Early warning: If you run this code, it will give an error.)
```py
import torch
from transformers import AutoModel
model = AutoModel.from_pretrained("google-bert/bert-base-uncased", device_map="auto")
attention_mask = tor... | closed | completed | false | 2 | [
"Feature request"
] | [] | 2026-06-06T04:00:17Z | 2026-06-10T16:03:00Z | 2026-06-10T16:03:00Z | CONTRIBUTOR | null | 20260610T180024Z | 2026-06-10T18:00:24Z | AbdiHaryadi | 48,970,896 | MDQ6VXNlcjQ4OTcwODk2 | User | false |
huggingface/transformers | 4,602,015,055 | I_kwDOCUB6oc8AAAABEk0tTw | 46,459 | https://github.com/huggingface/transformers/issues/46459 | https://api.github.com/repos/huggingface/transformers/issues/46459 | `dtype` is silently ignored when loading composite (multimodal) checkpoints through AutoModelForCausalLM (affects all Qwen3.5 repos) | ### System Info
- `transformers` version: 5.10.0.dev0 (main @ effde20942); also reproduced on the v5.9.0 release
- Platform: macOS-26.1-arm64-arm-64bit (Apple Silicon)
- Python version: 3.11.15
- PyTorch version (GPU?): 2.11.0 (False; MPS available, repro runs on CPU)
- Huggingface_hub version: 1.14.0
- Safetensors ve... | closed | completed | false | 4 | [] | [] | 2026-06-06T05:06:59Z | 2026-06-11T11:34:58Z | 2026-06-11T11:34:58Z | CONTRIBUTOR | null | 20260611T120117Z | 2026-06-11T12:01:17Z | qflen | 194,738,340 | U_kgDOC5t4pA | User | false |
huggingface/transformers | 4,603,789,150 | I_kwDOCUB6oc8AAAABEmg_Xg | 46,463 | https://github.com/huggingface/transformers/issues/46463 | https://api.github.com/repos/huggingface/transformers/issues/46463 | [Bug] Incorrect error message for T5 model family's decoder input validation | ### System Info
- `transformers` version: 5.5.1
- Platform: Windows-11-10.0.26200-SP0
- 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 (accelerator?): 2.6.0+cu124 (... | open | null | false | 3 | [
"bug"
] | [] | 2026-06-06T14:24:41Z | 2026-06-08T12:11:00Z | null | NONE | null | 20260608T180018Z | 2026-06-08T18:00:18Z | Reigenleif | 96,337,290 | U_kgDOBb39ig | User | false |
huggingface/transformers | 4,605,895,931 | I_kwDOCUB6oc8AAAABEohk-w | 46,468 | https://github.com/huggingface/transformers/issues/46468 | https://api.github.com/repos/huggingface/transformers/issues/46468 | [Bug] Official Nemotron example broken | ### System Info
transformers env fails with:
NameError: name 'CompletionCreateParamsStreaming' is not defined
but my pyproject if fairly simple (and python and cuda and packages are in here):
```
[project]
name = "nemotron-explorations"
version = "0.1.0"
description = "Add your description here"
readme = "README.md"... | open | null | false | 1 | [
"bug"
] | [] | 2026-06-07T02:23:21Z | 2026-06-07T13:23:59Z | null | NONE | null | 20260607T180020Z | 2026-06-07T18:00:20Z | Davidnet | 3,343,006 | MDQ6VXNlcjMzNDMwMDY= | User | false |
huggingface/transformers | 4,607,103,315 | I_kwDOCUB6oc8AAAABEprRUw | 46,471 | https://github.com/huggingface/transformers/issues/46471 | https://api.github.com/repos/huggingface/transformers/issues/46471 | New RAG framework: VORTEXRAG — 7-layer causal RAG achieving EM 74.8, Faithfulness 0.94 | Hi Transformers team!
Sharing **VORTEXRAG** — a new RAG framework built on top of HuggingFace Transformers.
VORTEXRAG uses:
- **all-mpnet-base-v2** (sentence-transformers) for the 768d semantic arm
- **DeBERTa-v3-small** CrossEncoder for the Faithfulness Verifier (FV) layer
- spaCy + PropBank for syntactic/causal arm... | open | null | false | 0 | [] | [] | 2026-06-07T11:14:19Z | 2026-06-07T11:14:19Z | null | NONE | null | 20260607T120016Z | 2026-06-07T12:00:16Z | vignesh2027 | 207,180,873 | U_kgDODFlUSQ | User | false |
huggingface/transformers | 4,610,291,353 | I_kwDOCUB6oc8AAAABEst2mQ | 46,489 | https://github.com/huggingface/transformers/issues/46489 | https://api.github.com/repos/huggingface/transformers/issues/46489 | DeepSeek-Coder v1 tokenizer produces incorrect output on transformers v5+ (gap in PR #44801's fix) | ### System Info
- `transformers` version: 5.10.2
- `tokenizers` version: 0.22.2
- Python version: 3.11.2
- Platform: Linux (Penn State HPC, RHEL 8.10)
- PyTorch version: not relevant (tokenizer-only repro)
### Who can help?
@ArthurZucker @itazap
### Information
- [ ] The official example scripts
- [ ] My own modif... | closed | completed | false | 5 | [
"bug"
] | [] | 2026-06-08T05:38:38Z | 2026-06-28T04:09:49Z | 2026-06-28T04:09:49Z | NONE | null | 20260628T060020Z | 2026-06-28T06:00:20Z | SuryanshSS1011 | 159,204,949 | U_kgDOCX1GVQ | User | false |
huggingface/transformers | 3,718,179,063 | I_kwDOCUB6oc7dnuj3 | 42,796 | https://github.com/huggingface/transformers/issues/42796 | https://api.github.com/repos/huggingface/transformers/issues/42796 | Ministral-3-8B-Instruct tokenizer doesn't handle BPE markers properly | ### System Info
```python
from transformers import AutoTokenizer, AutoModelForImageTextToText, AutoProcessor
import torch
base_model = "mistralai/Ministral-3-8B-Instruct-2512-BF16"
model = AutoModelForImageTextToText.from_pretrained(base_model, dtype=torch.bfloat16)
model = model.to("cuda:1")
tokenizer = AutoProcess... | closed | completed | false | 9 | [
"bug"
] | [] | 2025-12-11T07:35:35Z | 2026-06-08T06:23:31Z | 2026-01-08T10:49:33Z | NONE | null | 20260608T120020Z | 2026-06-08T12:00:20Z | yanancai | 23,553,208 | MDQ6VXNlcjIzNTUzMjA4 | User | true |
huggingface/transformers | 4,611,169,582 | I_kwDOCUB6oc8AAAABEtjdLg | 46,493 | https://github.com/huggingface/transformers/issues/46493 | https://api.github.com/repos/huggingface/transformers/issues/46493 | difference in performance of sam3 video model compared to original sam3 implementation | Hi,
i ran the original sam3 video model (from meta) and the transformers sam3 video model on the same video with the same text prompt.
I noticed that the original sam3 model detections were more accurate, and even that in the transformers' version some frames didn't have detections at all (whereas the original sam3 di... | open | null | false | 11 | [] | [] | 2026-06-08T08:17:48Z | 2026-06-10T11:48:57Z | null | NONE | null | 20260610T120053Z | 2026-06-10T12:00:53Z | iariav | 28,890,865 | MDQ6VXNlcjI4ODkwODY1 | User | false |
huggingface/transformers | 4,610,520,003 | I_kwDOCUB6oc8AAAABEs7zww | 46,491 | https://github.com/huggingface/transformers/issues/46491 | https://api.github.com/repos/huggingface/transformers/issues/46491 | CodeLlama tokenizer strips one leading space on encode→decode round-trip (regression vs v4) | ### System Info
- `transformers` version: 5.10.2
- `tokenizers` version: 0.22.2
- Python version: 3.11.2
- Platform: Linux (Penn State HPC, RHEL 8.10)
- PyTorch version: not relevant (tokenizer-only repro)
### Who can help?
@ArthurZucker @itazap
### Information
- [ ] The official example scripts
- [ ] My own modif... | open | null | false | 5 | [
"bug"
] | [] | 2026-06-08T06:26:02Z | 2026-06-12T05:38:20Z | null | NONE | null | 20260612T060053Z | 2026-06-12T06:00:53Z | SuryanshSS1011 | 159,204,949 | U_kgDOCX1GVQ | User | false |
huggingface/transformers | 2,083,730,860 | I_kwDOCUB6oc58MzWs | 28,530 | https://github.com/huggingface/transformers/issues/28530 | https://api.github.com/repos/huggingface/transformers/issues/28530 | Early stopping required metric_for_best_model, but did not find eval_f1 so early stopping is disabled | ### System Info
- `transformers` version: 4.35.2
- Platform: Linux-3.10.0-1160.49.1.el7.x86_64-x86_64-with-glibc2.17
- Python version: 3.8.18
- Huggingface_hub version: 0.19.4
- Safetensors version: 0.4.1
- Accelerate version: 0.25.0
- Accelerate config: not found
- PyTorch version (GPU?): 1.13.1+cu116 (Tr... | closed | completed | false | 17 | [
"Good Second Issue",
"bug"
] | [] | 2024-01-16T11:37:37Z | 2026-06-08T11:47:07Z | 2026-06-08T11:47:07Z | NONE | null | 20260608T120020Z | 2026-06-08T12:00:20Z | ManishChandra12 | 17,062,142 | MDQ6VXNlcjE3MDYyMTQy | User | false |
huggingface/transformers | 4,616,703,216 | I_kwDOCUB6oc8AAAABEy1M8A | 46,510 | https://github.com/huggingface/transformers/issues/46510 | https://api.github.com/repos/huggingface/transformers/issues/46510 | Your docs show how to use models but not what they cost | spam | closed | completed | false | 0 | [] | [] | 2026-06-08T21:16:27Z | 2026-06-09T11:32:27Z | 2026-06-09T11:31:50Z | NONE | null | 20260609T120115Z | 2026-06-09T12:01:15Z | agiulucom42-del | 246,366,963 | U_kgDODq9C8w | User | false |
huggingface/transformers | 4,615,395,797 | I_kwDOCUB6oc8AAAABExlZ1Q | 46,506 | https://github.com/huggingface/transformers/issues/46506 | https://api.github.com/repos/huggingface/transformers/issues/46506 | Different results for PPDocLayoutV3 on CPU and CUDA | ### System Info
PPDocLayoutV3 seems to have a numerical precision issue that snowballs into larger discrepancies. One cause of those precision issues is the function `_cached_generate_anchors`, which creates a binary mask. The mask may or may not contain the bottom-right boundary, depending on whether the chosen execu... | open | null | false | 9 | [
"bug"
] | [] | 2026-06-08T18:06:28Z | 2026-06-11T21:58:20Z | null | CONTRIBUTOR | null | 20260612T000041Z | 2026-06-12T00:00:41Z | 99991 | 18,725,165 | MDQ6VXNlcjE4NzI1MTY1 | User | false |
huggingface/transformers | 4,617,206,296 | I_kwDOCUB6oc8AAAABEzT6GA | 46,512 | https://github.com/huggingface/transformers/issues/46512 | https://api.github.com/repos/huggingface/transformers/issues/46512 | from_config(dtype=...) builds weights in the requested dtype but leaves model.config.dtype stale | ### System Info
- `transformers` version: 5.10.0.dev0 (also reproduces on current `main`)
- Platform: macOS-26.1-arm64-arm-64bit
- Python version: 3.11.15
- Huggingface_hub version: 1.14.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- PyTorch version (accelerator?): 2.11.0 (NA)
- Using distributed or par... | open | null | false | 2 | [] | [] | 2026-06-08T22:51:05Z | 2026-06-09T15:42:38Z | null | CONTRIBUTOR | null | 20260609T180023Z | 2026-06-09T18:00:23Z | qflen | 194,738,340 | U_kgDOC5t4pA | User | false |
huggingface/transformers | 4,618,790,131 | I_kwDOCUB6oc8AAAABE00k8w | 46,515 | https://github.com/huggingface/transformers/issues/46515 | https://api.github.com/repos/huggingface/transformers/issues/46515 | incompatibility with torch 2.4.1 | ### System Info
When I tested the latest version of transformers 5.10.2, I got the error: `AttributeError: module 'torch' has no attribute 'float8_e8m0fnu'`.
here is the full log:
```
rfdetr-1 | Traceback (most recent call last):
rfdetr-1 | File "/app/test.py", line 1, in <module>
rfdetr-1 | from rfdetr impo... | closed | completed | false | 2 | [
"bug"
] | [] | 2026-06-09T04:22:15Z | 2026-06-09T13:27:26Z | 2026-06-09T13:27:26Z | NONE | null | 20260609T180023Z | 2026-06-09T18:00:23Z | nickjyj | 46,984,040 | MDQ6VXNlcjQ2OTg0MDQw | User | false |
huggingface/transformers | 4,618,887,728 | I_kwDOCUB6oc8AAAABE06iMA | 46,516 | https://github.com/huggingface/transformers/issues/46516 | https://api.github.com/repos/huggingface/transformers/issues/46516 | [Continuous batching] Tracker for issues and feature requests | # Continuous batching issue and feature tracker
## Introduction
### Core idea
Continuous batching is the throughput-focused generation API for the `transformers` library, available through the `generate_batch` method.
It aims to provide a highly optimized generation API to a wide range of `transformers` models across... | open | null | false | 3 | [
"Performance",
"Inference"
] | [
"remi-or"
] | 2026-06-09T04:40:06Z | 2026-06-17T09:57:08Z | null | MEMBER | null | 20260617T120115Z | 2026-06-17T12:01:15Z | remi-or | 83,456,801 | MDQ6VXNlcjgzNDU2ODAx | User | false |
huggingface/transformers | 4,620,836,771 | I_kwDOCUB6oc8AAAABE2xfow | 46,518 | https://github.com/huggingface/transformers/issues/46518 | https://api.github.com/repos/huggingface/transformers/issues/46518 | Investigate training support for Sapiens2ForPoseEstimation when labels are provided | ### Feature request
While exploring the recently added `Sapiens2ForPoseEstimation` implementation, I noticed that the model's `forward()` method accepts `labels` (documented as heatmap ground truth with shape `(batch_size, num_keypoints, height, width)`), but currently raises `NotImplementedError("Training is not yet ... | closed | completed | false | 3 | [
"Feature request"
] | [] | 2026-06-09T09:43:45Z | 2026-06-25T09:50:29Z | 2026-06-25T09:50:29Z | CONTRIBUTOR | null | 20260625T120018Z | 2026-06-25T12:00:18Z | Sainava | 158,836,146 | U_kgDOCXelsg | User | false |
huggingface/transformers | 4,621,562,138 | I_kwDOCUB6oc8AAAABE3dxGg | 46,519 | https://github.com/huggingface/transformers/issues/46519 | https://api.github.com/repos/huggingface/transformers/issues/46519 | StopStringCriteria misses CJK stop strings on byte-level tokenizers when a character splits into byte-fragment tokens | On a byte-level BPE tokenizer (e.g. Qwen2.5), `StopStringCriteria` doesn't halt when the stop string's bytes are split across byte-fragment tokens, even though the stop string is the suffix of the decoded text. This hits CJK stop strings: a CJK character can fragment into tokens that each decode to U+FFFD in isolation,... | closed | completed | false | 5 | [] | [] | 2026-06-09T11:35:01Z | 2026-06-12T12:28:48Z | 2026-06-12T12:28:48Z | CONTRIBUTOR | null | 20260612T180035Z | 2026-06-12T18:00:35Z | Incheonkirin | 42,427,560 | MDQ6VXNlcjQyNDI3NTYw | User | false |
huggingface/transformers | 2,797,039,519 | I_kwDOCUB6oc6mt2-f | 35,766 | https://github.com/huggingface/transformers/issues/35766 | https://api.github.com/repos/huggingface/transformers/issues/35766 | Defining LLM Dataset types in Trainers or during Training Workflow | ### Feature request
Hi,
Just wondering if its possible to define a way to define the dataset input labels mapping argument in the transformers.Trainer or something similar?
I understand that the input data labels are dependent on the type of training loss function used and problem for example, the type of defining ... | closed | duplicate | false | 3 | [
"Feature request"
] | [] | 2025-01-18T15:37:12Z | 2026-06-09T18:45:12Z | 2026-06-09T18:44:26Z | NONE | null | 20260610T000024Z | 2026-06-10T00:00:24Z | whoamimi | 159,568,507 | U_kgDOCYLSew | User | false |
huggingface/transformers | 4,628,715,450 | I_kwDOCUB6oc8AAAABE-SXug | 46,531 | https://github.com/huggingface/transformers/issues/46531 | https://api.github.com/repos/huggingface/transformers/issues/46531 | [Gemma 4] Gemma4UnifiedForConditionalGeneration text-only inference produces degenerate output (token repetition collapse) | - `transformers`: latest `main` (`5.10.0.dev0`, tested 2026-06-08)
- `torch`: 2.7.0
- Platform: macOS, Apple M4 Pro (reproduced on **both MPS and CPU**)
- Python: 3.12
The collapse is **device-, dtype-, and attention-implementation-independent** — it reproduces on CPU with `float32` and `attn_implementation="eager"`, ... | closed | completed | false | 2 | [] | [] | 2026-06-10T06:39:36Z | 2026-06-11T08:48:10Z | 2026-06-11T08:48:10Z | NONE | null | 20260611T120117Z | 2026-06-11T12:01:17Z | synapticode-ai | 263,250,568 | U_kgDOD7DiiA | User | false |
huggingface/transformers | 4,630,423,057 | I_kwDOCUB6oc8AAAABE_6mEQ | 46,536 | https://github.com/huggingface/transformers/issues/46536 | https://api.github.com/repos/huggingface/transformers/issues/46536 | [Gemma4/Gemma3n] MultimodalEmbedder.forward missing dtype alignment that subclass override already applies | ### Summary
`Gemma4MultimodalEmbedder.forward` and `Gemma3nMultimodalEmbedder.forward` (on the `inputs_embeds` branch) feed activations directly into an `nn.Linear` projection without aligning the activation dtype with `embedding_projection.weight.dtype`. This is the only place in either model's vision/audio→LM bridge... | closed | completed | false | 3 | [] | [] | 2026-06-10T10:43:58Z | 2026-06-10T12:33:51Z | 2026-06-10T12:33:50Z | NONE | null | 20260610T180024Z | 2026-06-10T18:00:24Z | BiggieW | 88,428,807 | MDQ6VXNlcjg4NDI4ODA3 | User | false |
huggingface/transformers | 2,730,897,382 | I_kwDOCUB6oc6ixi_m | 35,185 | https://github.com/huggingface/transformers/issues/35185 | https://api.github.com/repos/huggingface/transformers/issues/35185 | QuantizedCache first token processing is counterintuitive / worse than in papers | ### System Info
transformers==4.46.3
torch==2.5.1
(though, it does not depend on library versions)
### Who can help?
@sunmarc @zucchini-nlp
### Information
- [X] The official example scripts
- [ ] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQ... | closed | completed | false | 5 | [
"Quantization",
"bug",
"Generation"
] | [] | 2024-12-10T18:34:31Z | 2026-06-10T20:13:07Z | 2025-01-17T21:42:45Z | NONE | null | 20260611T000026Z | 2026-06-11T00:00:26Z | goodevening13 | 91,018,099 | MDQ6VXNlcjkxMDE4MDk5 | User | false |
huggingface/transformers | 4,638,240,022 | I_kwDOCUB6oc8AAAABFHXtFg | 46,552 | https://github.com/huggingface/transformers/issues/46552 | https://api.github.com/repos/huggingface/transformers/issues/46552 | Wav2Vec2CTCTokenizer in v5 treats the word delimiter as a special token, which leaks into get_special_tokens_mask and convert_ids_to_tokens |
### System Info
- transformers 5.10.0.dev0 (current main), also reproduces on 5.9.0
- last good: 4.57.6
- Python 3.12, macOS arm64 (CPU only)
### Who can help?
@eustlb
### Reproduction
Runs offline in a few seconds:
```python
import json, tempfile, os
from transformers import Wav2Vec2CTCTokenizer
d = tempfile.m... | open | null | false | 1 | [] | [] | 2026-06-11T07:33:43Z | 2026-06-12T03:15:21Z | null | NONE | null | 20260612T060053Z | 2026-06-12T06:00:53Z | ishan-1010 | 98,383,932 | U_kgDOBd04PA | User | false |
huggingface/transformers | 4,638,227,188 | I_kwDOCUB6oc8AAAABFHW69A | 46,551 | https://github.com/huggingface/transformers/issues/46551 | https://api.github.com/repos/huggingface/transformers/issues/46551 | Wav2Vec2CTCTokenizer in v5 treats the word delimiter as a special token, which leaks into get_special_tokens_mask and convert_ids_to_tokens |
### System Info
- transformers 5.10.0.dev0 (current main), also reproduces on 5.9.0
- last good: 4.57.6
- Python 3.12, macOS arm64 (CPU only)
### Who can help?
@eustlb
### Reproduction
Runs offline in a few seconds:
```python
import json, tempfile, os
from transformers import Wav2Vec2CTCTokenizer
d = tempfile.m... | closed | completed | false | 1 | [] | [] | 2026-06-11T07:31:34Z | 2026-06-11T07:33:56Z | 2026-06-11T07:33:56Z | NONE | null | 20260611T120117Z | 2026-06-11T12:01:17Z | SHAI-ishan-katoch | 198,710,568 | U_kgDOC9gVKA | User | false |
huggingface/transformers | 4,639,514,443 | I_kwDOCUB6oc8AAAABFIlfSw | 46,561 | https://github.com/huggingface/transformers/issues/46561 | https://api.github.com/repos/huggingface/transformers/issues/46561 | serve: Gemma 4 non-thinking responses returned as reasoning_content with empty content | ## System Info
- `transformers` main (`acc2cda7d9`), `transformers serve`
- Model: `google/gemma-4-31B-it`
- Linux, Python 3.x, CUDA
## Who can help?
Serving / CLI maintainers
## Reproduction
1. Start the server and send a non-streaming chat completion with a thinking-disabled prompt:
```bash
transformers serve
c... | closed | completed | false | 1 | [] | [] | 2026-06-11T10:24:08Z | 2026-06-30T06:24:11Z | 2026-06-30T06:24:11Z | NONE | null | 20260630T120021Z | 2026-06-30T12:00:21Z | b11015006 | 113,914,140 | U_kgDOBsoxHA | User | false |
huggingface/transformers | 4,639,409,095 | I_kwDOCUB6oc8AAAABFIfDxw | 46,559 | https://github.com/huggingface/transformers/issues/46559 | https://api.github.com/repos/huggingface/transformers/issues/46559 | `apply_chat_template(return_assistant_tokens_mask=True)` returns an empty assistant mask for VLM processors when an image is present | ### System Info
- `transformers` version: 5.10.1
- Platform: Linux
- Python: 3.13
@zucchini-nlp
### Description
For a multimodal processor, `apply_chat_template(..., return_assistant_tokens_mask=True)` produces a correct assistant mask for a text-only conversation, but an **all-zero (empty)** mask as soon as a real... | closed | duplicate | false | 2 | [] | [] | 2026-06-11T10:08:39Z | 2026-06-11T11:42:34Z | 2026-06-11T10:11:10Z | MEMBER | null | 20260611T120117Z | 2026-06-11T12:01:17Z | qgallouedec | 45,557,362 | MDQ6VXNlcjQ1NTU3MzYy | User | false |
huggingface/transformers | 4,642,220,863 | I_kwDOCUB6oc8AAAABFLKrPw | 46,566 | https://github.com/huggingface/transformers/issues/46566 | https://api.github.com/repos/huggingface/transformers/issues/46566 | DiffusionGemma weights remain on 'meta' device with device_map="auto" | Hello Hugging Face team,
Thank you for your amazing work on the `transformers` library. I encountered an issue when trying to run the new `DiffusionGemma` model using `device_map="auto"`:
```python
from transformers import DiffusionGemmaForBlockDiffusion, AutoProcessor
MODEL_ID = "google/diffusiongemma-26B-A4B-it"
... | closed | completed | false | 8 | [] | [] | 2026-06-11T16:08:46Z | 2026-06-17T09:59:42Z | 2026-06-17T09:59:42Z | CONTRIBUTOR | null | 20260617T120115Z | 2026-06-17T12:01:15Z | HuangBugWei | 67,520,151 | MDQ6VXNlcjY3NTIwMTUx | User | false |
huggingface/transformers | 4,643,077,274 | I_kwDOCUB6oc8AAAABFL-8mg | 46,570 | https://github.com/huggingface/transformers/issues/46570 | https://api.github.com/repos/huggingface/transformers/issues/46570 | AutoModel loading fails with ROCm when model has custom CUDA kernels | Models with custom CUDA extensions (e.g., flash-attention) fail to load on ROCm. The `is_available()` check passes but the actual kernel launch fails.
Expected: graceful fallback to standard attention.
Actual: RuntimeError at inference time.
ROCm 6.0, MI210 | open | null | false | 1 | [] | [] | 2026-06-11T18:16:24Z | 2026-06-12T10:41:47Z | null | NONE | null | 20260612T120034Z | 2026-06-12T12:00:34Z | hakuba-pamu | 290,990,455 | U_kgDOEVgpdw | User | false |
huggingface/transformers | 3,256,417,423 | I_kwDOCUB6oc7CGPyP | 39,607 | https://github.com/huggingface/transformers/issues/39607 | https://api.github.com/repos/huggingface/transformers/issues/39607 | ImageClassificationPipeline preprocess should accept numpy/tensor arrays | ### Feature request
Currently, `ImageClassificationPipeline` expects a PIL image or a string pointing to a URL.
This makes using existing datasets (e.g. from `torchvision`) a bit more difficult in the generic case. A non-transformers model would work with torch tensors, but not necessarily with a PIL image, whereas `I... | closed | completed | false | 4 | [
"Feature request"
] | [] | 2025-07-23T13:30:32Z | 2026-06-16T12:37:28Z | 2026-06-16T12:37:28Z | NONE | null | 20260616T180029Z | 2026-06-16T18:00:29Z | idantene | 12,184,618 | MDQ6VXNlcjEyMTg0NjE4 | User | false |
huggingface/transformers | 3,479,461,065 | I_kwDOCUB6oc7PZFzJ | 41,306 | https://github.com/huggingface/transformers/issues/41306 | https://api.github.com/repos/huggingface/transformers/issues/41306 | TiledFusedLogitsLoss (LigerFusedLinearCrossEntropy, FLCE) Flag in `from_pretrained()` | ### Feature request
Add a native `tiled_fused_logits_loss` flag in `from_pretrained()` that enables memory-efficient cross entropy computation across different backends.
Example API:
```python
from_pretrained(..., tiled_fused_logits_loss=Union[bool, List["auto", "liger-kernel", "deepspeed"])
```
Supported options co... | open | null | false | 13 | [
"Feature request"
] | [] | 2025-10-03T01:36:44Z | 2026-06-12T00:11:13Z | null | CONTRIBUTOR | null | 20260612T060053Z | 2026-06-12T06:00:53Z | Tcc0403 | 76,503,978 | MDQ6VXNlcjc2NTAzOTc4 | User | false |
huggingface/transformers | 4,645,816,402 | I_kwDOCUB6oc8AAAABFOmIUg | 46,577 | https://github.com/huggingface/transformers/issues/46577 | https://api.github.com/repos/huggingface/transformers/issues/46577 | Add RTDetrForSegmentation | ### Feature request
Add an `RTDetrForSegmentation` and `RTDetrV2ForSegmentation` class that extends the object detection model class with a mask prediction head, similar to `DetrForSegmentation`.
### Motivation
RT-DETR in the transformers library currently only supports bounding box predictions through `RTDetrForObj... | open | null | false | 1 | [
"Feature request"
] | [] | 2026-06-12T02:49:37Z | 2026-06-12T10:45:59Z | null | NONE | null | 20260612T120034Z | 2026-06-12T12:00:34Z | ranykamel | 25,268,334 | MDQ6VXNlcjI1MjY4MzM0 | User | false |
huggingface/transformers | 3,201,255,977 | I_kwDOCUB6oc6-z0op | 39,215 | https://github.com/huggingface/transformers/issues/39215 | https://api.github.com/repos/huggingface/transformers/issues/39215 | _load_rng_state after get_batch_samples may break training reproducibility when dataloader has random operations |
### Reproduction
The current implementation in the `Trainer`'s `_inner_training_loop` for resuming from a checkpoint calls `_load_rng_state` *after* fetching the data batch with `get_batch_samples`. This logic appears to be designed to handle the complexities of `skip_first_batches` and multi-worker dataloading.
... | open | reopened | false | 6 | [
"bug"
] | [] | 2025-07-04T04:14:46Z | 2026-06-16T01:25:40Z | null | CONTRIBUTOR | null | 20260616T060022Z | 2026-06-16T06:00:22Z | rangehow | 88,258,534 | MDQ6VXNlcjg4MjU4NTM0 | User | false |
huggingface/transformers | 4,648,944,777 | I_kwDOCUB6oc8AAAABFRlEiQ | 46,597 | https://github.com/huggingface/transformers/issues/46597 | https://api.github.com/repos/huggingface/transformers/issues/46597 | [`docs`] PULL_REQUEST_TEMPLATE.md has broken link to non-existent CONTRIBUTING.md section | Hello!
## Details
This line in the PULL_REQUEST_TEMPLATE.md links to https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#create-a-pull-request, but this section doesn't exist in `CONTRIBUTING.md` (anymore?):
https://github.com/huggingface/transformers/blob/a14eae2b54c19cb427c919a99c75db07afbeb7a0/.g... | closed | completed | false | 1 | [] | [] | 2026-06-12T11:47:18Z | 2026-06-15T15:43:54Z | 2026-06-15T15:43:54Z | MEMBER | null | 20260615T180022Z | 2026-06-15T18:00:22Z | tomaarsen | 37,621,491 | MDQ6VXNlcjM3NjIxNDkx | User | false |
huggingface/transformers | 4,648,995,061 | I_kwDOCUB6oc8AAAABFRoI9Q | 46,598 | https://github.com/huggingface/transformers/issues/46598 | https://api.github.com/repos/huggingface/transformers/issues/46598 | MCP to help OSS devs understand huggingface/transformers | Hey huggingface/transformers folks,
transformers is a big codebase, and I noticed newcomers are mostly confused about where to start. Digging through codebase through any AI agent will burn a lot of tokens for simplest task as it uses complete repo context.
I maintain [bytebell/open-ir](https://github.com/ByteBell/op... | open | null | false | 0 | [] | [] | 2026-06-12T11:54:47Z | 2026-06-12T11:54:47Z | null | NONE | null | 20260612T120034Z | 2026-06-12T12:00:34Z | ankitzm | 66,105,983 | MDQ6VXNlcjY2MTA1OTgz | User | false |
huggingface/transformers | 4,652,721,548 | I_kwDOCUB6oc8AAAABFVLljA | 46,614 | https://github.com/huggingface/transformers/issues/46614 | https://api.github.com/repos/huggingface/transformers/issues/46614 | Wav2Vec2PhonemeCTCTokenizer.phonemize keeps the previous language's espeak backend after switching back (stale phonemizer_lang) | ### System Info
transformers 4.57.6, phonemizer 3.3.0, espeak-ng via espeakng-loader 0.2.4,
Python 3.10, Linux.
### 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/SQ... | open | null | false | 3 | [
"bug"
] | [] | 2026-06-12T21:26:47Z | 2026-06-16T12:36:48Z | null | NONE | null | 20260616T180029Z | 2026-06-16T18:00:29Z | BlueCollarChris | 13,341,206 | MDQ6VXNlcjEzMzQxMjA2 | User | false |
huggingface/transformers | 4,652,445,295 | I_kwDOCUB6oc8AAAABFU6ubw | 46,612 | https://github.com/huggingface/transformers/issues/46612 | https://api.github.com/repos/huggingface/transformers/issues/46612 | Beam search cache reorder silently skipped for Mamba, XLNet, RWKV, and Reformer models (wrong generation output) | ### System Info
transformers version: main/latest
Platform: all
Python version: 3.10+
### Who can help?
@Cyrilvallez
### Information
- [ ] The official example scripts
- [ ] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [ ] My own task ... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-06-12T20:36:29Z | 2026-06-24T03:13:58Z | 2026-06-24T03:13:58Z | NONE | null | 20260624T060017Z | 2026-06-24T06:00:17Z | wi5nuu | 170,215,451 | U_kgDOCiVIGw | User | false |
huggingface/transformers | 4,654,404,361 | I_kwDOCUB6oc8AAAABFWyTCQ | 46,617 | https://github.com/huggingface/transformers/issues/46617 | https://api.github.com/repos/huggingface/transformers/issues/46617 | FileNotFoundError when loading a trust_remote_code model from cache | ### System Info
transformer version: 5.10.1
vllm 0.22
### Who can help?
@Cyrilvallez
### Information
- [ ] The official example scripts
- [ ] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [ ] My own task or dataset (give details below)
... | open | null | false | 2 | [
"bug"
] | [] | 2026-06-13T04:53:55Z | 2026-06-14T16:36:34Z | null | NONE | null | 20260614T180027Z | 2026-06-14T18:00:27Z | ldkhang1201 | 112,814,283 | U_kgDOBrloyw | User | false |
huggingface/transformers | 4,654,824,030 | I_kwDOCUB6oc8AAAABFXL6Xg | 46,620 | https://github.com/huggingface/transformers/issues/46620 | https://api.github.com/repos/huggingface/transformers/issues/46620 | transformers 5.x: from_pretrained silently returns random-init weights for custom PreTrainedModel subclasses (clean loading_info, no warning) | ### System Info
- transformers: 5.12.0
- torch: 2.12.0+cpu
- Python: 3.11
- Reproduces in a clean venv with only `transformers`, `torch`, `safetensors` installed (OS-independent).
### Description
A custom (out-of-tree) `PreTrainedModel` subclass saved with `save_pretrained()` and reloaded with `from_pretrained()` co... | closed | completed | false | 5 | [] | [] | 2026-06-13T07:58:47Z | 2026-06-22T15:38:53Z | 2026-06-22T15:38:53Z | NONE | null | 20260622T180028Z | 2026-06-22T18:00:28Z | gkriegspeedbay | 269,034,776 | U_kgDOEAklGA | User | false |
huggingface/transformers | 4,656,501,838 | I_kwDOCUB6oc8AAAABFYyUTg | 46,629 | https://github.com/huggingface/transformers/issues/46629 | https://api.github.com/repos/huggingface/transformers/issues/46629 | BUG? transformers version '5.12.0' gemma-4 generate DynamicSlidingWindowLayer |
Hi, I'm generating text with MTP but I'm getting the DynamicSlidingWindowLayer error.
If I disable MTP, it generates the text normally (# assistant_model=assistant_model,)
transformers version '5.12.0'
torch version
CODE: '2.11.0+cu128'
GPU: RTX 6000 Ada
```py
'''
import transformers, torch
transformers.__version__... | open | null | false | 1 | [] | [] | 2026-06-13T18:27:11Z | 2026-06-13T19:04:40Z | null | NONE | null | 20260614T000126Z | 2026-06-14T00:01:26Z | NickyDark1 | 123,802,672 | U_kgDOB2EUMA | User | false |
huggingface/transformers | 4,658,135,633 | I_kwDOCUB6oc8AAAABFaWCUQ | 46,640 | https://github.com/huggingface/transformers/issues/46640 | https://api.github.com/repos/huggingface/transformers/issues/46640 | Flash Attention 2 issue | ---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
[/tmp/ipykernel_1434/3174974862.py](https://localhost:8080/#) in <cell line: 0>()
2 import torch
3
----> 4 model = AutoModel.from_pretrained(
5 "... | open | null | false | 4 | [] | [] | 2026-06-14T06:19:04Z | 2026-06-16T06:39:23Z | null | NONE | null | 20260616T120046Z | 2026-06-16T12:00:46Z | srishtiimp | 252,687,560 | U_kgDODw-0yA | User | false |
huggingface/transformers | 2,066,557,346 | I_kwDOCUB6oc57LSmi | 28,348 | https://github.com/huggingface/transformers/issues/28348 | https://api.github.com/repos/huggingface/transformers/issues/28348 | Add flash attention 2.0 support for GPT2LMHeadModel | ```
model = AutoModelForCausalLM.from_pretrained(
my_GPT2LMHeadModel_checkpoint,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
```
throws the following error:
```
Error loading Flash_Model_2: GPT2LMHeadModel does not support Flash Attention 2.0 yet. Pl... | closed | completed | false | 2 | [
"Good Second Issue",
"Feature request"
] | [] | 2024-01-05T00:39:27Z | 2026-06-15T12:37:14Z | 2026-06-15T12:37:14Z | NONE | null | 20260615T180022Z | 2026-06-15T18:00:22Z | brresnic | 6,865,869 | MDQ6VXNlcjY4NjU4Njk= | User | false |
huggingface/transformers | 4,660,423,250 | I_kwDOCUB6oc8AAAABFchqUg | 46,644 | https://github.com/huggingface/transformers/issues/46644 | https://api.github.com/repos/huggingface/transformers/issues/46644 | Consider reverting CLIP refactor | ### System Info
Please consider reverting https://github.com/huggingface/transformers/pull/44431
Here is detailed information what problems it causes:
https://github.com/Nerogar/OneTrainer/pull/1506
Besides the one linked, it prevents multiple projects from upgrading or requires workarounds/conversion to match exist... | open | null | false | 4 | [
"bug"
] | [] | 2026-06-14T20:01:08Z | 2026-06-22T09:00:18Z | null | NONE | null | 20260622T120018Z | 2026-06-22T12:00:18Z | dxqb | 183,307,934 | U_kgDOCu0Ong | User | false |
huggingface/transformers | 4,661,505,910 | I_kwDOCUB6oc8AAAABFdjvdg | 46,645 | https://github.com/huggingface/transformers/issues/46645 | https://api.github.com/repos/huggingface/transformers/issues/46645 | Add SWA+sink option to Granite models | ### Model description
For upcoming Granite model release (4.5) we would like to add support for sliding window attention with a sink token. I think it may be easier to add as a flag in existing models, rather than adding more Granite model classes (granitemoesharedswa seems gratuitous), though we'll need to ensure no ... | open | null | false | 2 | [
"New model"
] | [] | 2026-06-15T02:09:07Z | 2026-06-16T22:56:02Z | null | CONTRIBUTOR | null | 20260617T000107Z | 2026-06-17T00:01:07Z | daviswer | 9,604,893 | MDQ6VXNlcjk2MDQ4OTM= | User | false |
huggingface/transformers | 4,662,183,981 | I_kwDOCUB6oc8AAAABFeNILQ | 46,650 | https://github.com/huggingface/transformers/issues/46650 | https://api.github.com/repos/huggingface/transformers/issues/46650 | `save_pretrained` corrupts keys (`encoder.encoder.*`) for trust_remote_code / from_config models via non-idempotent reverse rename | ### System Info
```
- `transformers` version: 5.5.3
- Platform: Linux-5.15.0-100-generic-x86_64-with-glibc2.35
- Python version: 3.11.12
- Huggingface_hub version: 1.6.0
- PyTorch version (accelerator?): 2.10.0+cu128 (CUDA)
- Using GPU in script?: No (pure key-renaming logic)
```
Also reproduces on `main` (`5.13.0.de... | closed | completed | false | 0 | [] | [] | 2026-06-15T04:58:50Z | 2026-06-29T04:41:59Z | 2026-06-29T04:41:59Z | CONTRIBUTOR | null | 20260629T060016Z | 2026-06-29T06:00:16Z | Bluear7878 | 200,191,325 | U_kgDOC-6tXQ | User | false |
huggingface/transformers | 4,663,972,250 | I_kwDOCUB6oc8AAAABFf6Rmg | 46,656 | https://github.com/huggingface/transformers/issues/46656 | https://api.github.com/repos/huggingface/transformers/issues/46656 | Inconsistent type annotation in `EvalPrediction` | https://github.com/huggingface/transformers/blob/052e652d6d53c2b26ffde87e039b723949a53493/src/transformers/trainer.py#L4211
Hi there,
I am experimenting with the Trainer and noticed a potentially inconsistent type annotation for the fields in `EvalPrediction`. In particular, when the `batch_eval_metrics` is True the ... | open | null | false | 4 | [] | [] | 2026-06-15T09:54:50Z | 2026-06-25T18:57:31Z | null | CONTRIBUTOR | null | 20260626T000017Z | 2026-06-26T00:00:17Z | pietrolesci | 61,748,653 | MDQ6VXNlcjYxNzQ4NjUz | User | false |
huggingface/transformers | 4,663,967,394 | I_kwDOCUB6oc8AAAABFf5-og | 46,655 | https://github.com/huggingface/transformers/issues/46655 | https://api.github.com/repos/huggingface/transformers/issues/46655 | SmolVLM: Error due to incorrect padding calculation in video processor | ### System Info
transformers: 5.2.0
torch: 2.10
### Who can help?
@zucchini-nlp
### 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"
] | [] | 2026-06-15T09:54:17Z | 2026-06-16T09:38:45Z | 2026-06-16T09:38:45Z | NONE | null | 20260616T120046Z | 2026-06-16T12:00:46Z | andreasgoulas | 1,699,783 | MDQ6VXNlcjE2OTk3ODM= | User | false |
huggingface/transformers | 4,663,988,409 | I_kwDOCUB6oc8AAAABFf7QuQ | 46,657 | https://github.com/huggingface/transformers/issues/46657 | https://api.github.com/repos/huggingface/transformers/issues/46657 | Size of tensors mismatch when training object detection model | ### System Info
Unfortunately transformers env fails with the following error: NameError: name 'CompletionCreateParamsStreaming' is not defined. Please find manuel configuration instead:
System: Linux 6.1.0-40-amd64 (Docker)
Python version: 3.12.13
transformers==5.12.0
### Who can help?
_No response_
### Informatio... | closed | completed | false | 7 | [
"bug"
] | [] | 2026-06-15T09:57:03Z | 2026-06-19T07:33:04Z | 2026-06-19T07:33:03Z | NONE | null | 20260619T120025Z | 2026-06-19T12:00:25Z | Rayndell | 40,670,939 | MDQ6VXNlcjQwNjcwOTM5 | User | false |
huggingface/transformers | 4,664,385,377 | I_kwDOCUB6oc8AAAABFgTfYQ | 46,661 | https://github.com/huggingface/transformers/issues/46661 | https://api.github.com/repos/huggingface/transformers/issues/46661 | AutoModel.from_pretrained raises AttributeError: model_class has no attribute 'config_class' for remote-code models | ### System Info
- `transformers` version: 5.13.0.dev0
- Platform: Linux-6.6.122+-x86_64-with-glibc2.35
- Python version: 3.12.13
- Huggingface_hub version: 1.18.0
- Safetensors version: 0.8.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?)... | closed | completed | false | 5 | [
"bug"
] | [] | 2026-06-15T10:52:55Z | 2026-06-15T15:23:52Z | 2026-06-15T15:23:52Z | CONTRIBUTOR | null | 20260615T180022Z | 2026-06-15T18:00:22Z | atharv1945 | 138,139,938 | U_kgDOCDvZIg | User | false |
huggingface/transformers | 4,664,781,728 | I_kwDOCUB6oc8AAAABFgrroA | 46,664 | https://github.com/huggingface/transformers/issues/46664 | https://api.github.com/repos/huggingface/transformers/issues/46664 | SmolVLM: Inconsistent meaning of num_frames in video processor | ### System Info
transformers: 5.12.0
### Who can help?
@zucchini-nlp
### Information
- [ ] The official example scripts
- [ ] My own modified scripts
### Tasks
- [ ] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [ ] My own task or dataset (give details below)
### Reproductio... | closed | not_planned | false | 2 | [
"bug"
] | [] | 2026-06-15T11:49:51Z | 2026-06-15T12:26:48Z | 2026-06-15T12:26:48Z | NONE | null | 20260615T180022Z | 2026-06-15T18:00:22Z | andreasgoulas | 1,699,783 | MDQ6VXNlcjE2OTk3ODM= | User | false |
huggingface/transformers | 4,663,450,014 | I_kwDOCUB6oc8AAAABFfaZng | 46,653 | https://github.com/huggingface/transformers/issues/46653 | https://api.github.com/repos/huggingface/transformers/issues/46653 | Mirko Privitera | ### Model description
30 settembre 1990 Palermo ma abito a sustinenza vicino casaleone provincia di Verona 3247929218
### Open source status
- [x] The model implementation is available
- [x] The model weights are available
### Provide useful links for the implementation
_No response_ | closed | completed | false | 1 | [
"New model"
] | [] | 2026-06-15T08:42:07Z | 2026-06-15T11:58:55Z | 2026-06-15T11:58:55Z | NONE | null | 20260615T120125Z | 2026-06-15T12:01:25Z | mirko772 | 207,275,767 | U_kgDODFrG9w | User | false |
huggingface/transformers | 3,304,882,234 | I_kwDOCUB6oc7E_IA6 | 40,042 | https://github.com/huggingface/transformers/issues/40042 | https://api.github.com/repos/huggingface/transformers/issues/40042 | Support loading glm4moe GGUF | ### Feature request
Currently, GGUF versions of GLM-4.5 series MoE models raises "GGUF model with architecture glm4moe is not supported yet" error.
### Motivation
GLM-4.5 series MoE GGUF models will successfully run. That is a fantastic model ...
### Your contribution
nope ... | open | null | false | 7 | [
"Good Second Issue",
"Feature request"
] | [] | 2025-08-08T18:27:44Z | 2026-07-03T17:40:24Z | null | NONE | null | 20260703T180022Z | 2026-07-03T18:00:22Z | adonishong | 9,166,212 | MDQ6VXNlcjkxNjYyMTI= | User | false |
huggingface/transformers | 2,146,978,349 | I_kwDOCUB6oc5_-Eot | 29,177 | https://github.com/huggingface/transformers/issues/29177 | https://api.github.com/repos/huggingface/transformers/issues/29177 | OSError: You are trying to access a gated repo. | ### Model description
I have submit access request to through huggingface and granted me access but not able to run model on inference.
``` Python
import torch
from torch import cuda, bfloat16
import transformers
model_id = 'google/gemma-7b'
device = f'cuda:{cuda.current_device()}' if cuda.is_available() el... | closed | completed | false | 30 | [
"New model"
] | [] | 2024-02-21T15:03:42Z | 2026-06-15T13:02:59Z | 2024-02-22T12:30:17Z | NONE | null | 20260615T180022Z | 2026-06-15T18:00:22Z | KaifAhmad1 | 98,801,504 | U_kgDOBeOXYA | User | false |
huggingface/transformers | 4,664,902,147 | I_kwDOCUB6oc8AAAABFgzCAw | 46,665 | https://github.com/huggingface/transformers/issues/46665 | https://api.github.com/repos/huggingface/transformers/issues/46665 | Inconsistent behaviour when using `from_pretrained` inside Python classes | ### System Info
- `transformers` version: 5.4.0
- Platform: Linux-6.17.0-35-generic-x86_64-with-glibc2.39
- Python version: 3.14.3
- Huggingface_hub version: 1.7.1
- Safetensors version: 0.7.0
- Accelerate version: 1.14.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerat... | closed | completed | false | 4 | [
"bug"
] | [] | 2026-06-15T12:06:39Z | 2026-06-16T08:39:55Z | 2026-06-15T15:15:09Z | NONE | null | 20260616T120046Z | 2026-06-16T12:00:46Z | I-Absolutely-Am-Not-A-Robot | 60,075,386 | MDQ6VXNlcjYwMDc1Mzg2 | User | false |
huggingface/transformers | 2,374,347,292 | I_kwDOCUB6oc6Nhaoc | 31,627 | https://github.com/huggingface/transformers/issues/31627 | https://api.github.com/repos/huggingface/transformers/issues/31627 | Tokenizer discard data that exceed max_length | ### Feature request
When use tokenizer, it truncate data to max_length, but can't just discard the data.
### Motivation
Sometimes we want the sentence to be complete
### Your contribution
No | closed | completed | false | 5 | [
"Core: Tokenization",
"Feature request"
] | [] | 2024-06-26T05:50:49Z | 2026-06-16T12:19:24Z | 2026-06-16T12:19:24Z | NONE | null | 20260616T180029Z | 2026-06-16T18:00:29Z | fengyunflya | 7,424,485 | MDQ6VXNlcjc0MjQ0ODU= | User | false |
huggingface/transformers | 4,670,840,631 | I_kwDOCUB6oc8AAAABFmdfNw | 46,682 | https://github.com/huggingface/transformers/issues/46682 | https://api.github.com/repos/huggingface/transformers/issues/46682 | [Flex Attention] `Tensor` q_offset will cause wrong mask shape | ### System Info
```
torch 2.14.0.dev20260613+cu126
torchao 0.18.0.dev20260615+cu126
transformers 5.12.1
triton 3.7.1+git5d6048aa
```
### Who can help?
@ArthurZucker @CyrilVallez
### Information
- [ ] The official example scripts
- [x] My own modifie... | closed | completed | false | 4 | [
"bug"
] | [] | 2026-06-16T03:41:19Z | 2026-06-23T11:00:16Z | 2026-06-23T11:00:16Z | NONE | null | 20260623T120018Z | 2026-06-23T12:00:18Z | hoshibara | 108,672,484 | U_kgDOBno15A | User | false |
huggingface/transformers | 4,672,286,086 | I_kwDOCUB6oc8AAAABFn1thg | 46,688 | https://github.com/huggingface/transformers/issues/46688 | https://api.github.com/repos/huggingface/transformers/issues/46688 | [Qwen2VL] Image processor backend refactor (#43514) introduces pixel_values precision drift, causing bbox coordinate shift | ## Title
[Qwen2VL] Image processor backend refactor (#43514) introduces pixel_values precision drift, causing bbox coordinate shift
## Labels
bug, Vision
## Body
### System Info
- `transformers` 4.57.1 vs 5.5.4
- `torchvision` 0.22.0
- `torch` 2.7.0
- Model: `Qwen3-VL-32B-Instruct`
- GPU: NVIDIA H20 (baseline) vs... | open | null | false | 3 | [] | [] | 2026-06-16T08:05:30Z | 2026-06-16T09:18:51Z | null | NONE | null | 20260616T120046Z | 2026-06-16T12:00:46Z | chenxi971215 | 43,957,823 | MDQ6VXNlcjQzOTU3ODIz | User | false |
huggingface/transformers | 4,674,541,488 | I_kwDOCUB6oc8AAAABFp_XsA | 46,692 | https://github.com/huggingface/transformers/issues/46692 | https://api.github.com/repos/huggingface/transformers/issues/46692 | Out of memory after some epochs while training RT-DETR v2 model | ### System Info
Unfortunately transformers env fails with the following error: NameError: name 'CompletionCreateParamsStreaming' is not defined. Please find manuel configuration instead:
System: Linux 6.1.0-40-amd64 (Docker)
Python version: 3.12.13
transformers==5.12.0
### Who can help?
_No response_
### Informatio... | open | null | false | 10 | [
"bug"
] | [] | 2026-06-16T13:18:42Z | 2026-06-24T18:10:50Z | null | NONE | null | 20260625T000027Z | 2026-06-25T00:00:27Z | Rayndell | 40,670,939 | MDQ6VXNlcjQwNjcwOTM5 | User | false |
huggingface/transformers | 4,675,063,993 | I_kwDOCUB6oc8AAAABFqfQuQ | 46,693 | https://github.com/huggingface/transformers/issues/46693 | https://api.github.com/repos/huggingface/transformers/issues/46693 | Performance regression with FA2 and long-context generation after 5.2.0 | ### System Info
transformers: 5.2.0 vs 5.4.0 (I could not test 5.3.0 due to a bug)
PyTorch: 2.12
flash-attn: 2.8.3
CUDA: 13.2
Python: 3.12
cudnn: 9.20.0.48
### Who can help?
@zucchini-nlp
@Cyrilvallez
### Information
- [ ] The official example scripts
- [x] My own modified scripts
### Tasks
- [ ] An officially... | open | null | false | 15 | [
"bug"
] | [] | 2026-06-16T14:21:31Z | 2026-07-02T07:14:52Z | null | NONE | null | 20260702T120025Z | 2026-07-02T12:00:25Z | andreasgoulas | 1,699,783 | MDQ6VXNlcjE2OTk3ODM= | User | false |
huggingface/transformers | 3,304,119,793 | I_kwDOCUB6oc7E8N3x | 40,032 | https://github.com/huggingface/transformers/issues/40032 | https://api.github.com/repos/huggingface/transformers/issues/40032 | Add Padding Strategy to DataCollatorForLanguageModeling | ### Feature request
Add the ability to specify a padding strategy when using `DataCollatorForLanguageModeling`
### Motivation
This is a minor QOL enhancement that makes the collator more consistent with others in the library. The main use case would probably be padding to max length to make memory usage more stable ... | closed | completed | false | 1 | [
"Feature request"
] | [] | 2025-08-08T13:46:00Z | 2026-06-17T11:50:17Z | 2026-06-17T11:50:17Z | CONTRIBUTOR | null | 20260617T120115Z | 2026-06-17T12:01:15Z | rjgleaton | 70,818,603 | MDQ6VXNlcjcwODE4NjAz | User | false |
huggingface/transformers | 4,677,218,640 | I_kwDOCUB6oc8AAAABFsixUA | 46,697 | https://github.com/huggingface/transformers/issues/46697 | https://api.github.com/repos/huggingface/transformers/issues/46697 | Swin2SR image processor over-pads images when dimensions are already divisible by size_divisor | ## System Info
```
- `transformers` version: 4.57.6
- Platform: macOS-26.5.1-arm64-arm-64bit
- Python version: 3.9.6
- PyTorch version: 2.8.0
- CUDA available: False
```
## Who can help?
@yonigozlan @molbap (vision models)
## Information
- [ ] The official example scripts
- [x] My own modified scripts
## Tasks
-... | open | null | false | 1 | [] | [] | 2026-06-16T19:21:58Z | 2026-06-17T15:11:47Z | null | NONE | null | 20260617T180027Z | 2026-06-17T18:00:27Z | arnavkewalram | 154,707,184 | U_kgDOCTik8A | User | false |
huggingface/transformers | 4,679,174,766 | I_kwDOCUB6oc8AAAABFuaKbg | 46,701 | https://github.com/huggingface/transformers/issues/46701 | https://api.github.com/repos/huggingface/transformers/issues/46701 | [DiffusionGemmaGenerationOutput] image-text-to-text pipeline does not work | ### System Info
transformers == 5.12.1
### Who can help?
@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)
### Reproduct... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-06-17T01:22:38Z | 2026-06-23T09:54:48Z | 2026-06-23T09:54:48Z | CONTRIBUTOR | null | 20260623T120018Z | 2026-06-23T12:00:18Z | KoichiYasuoka | 15,098,598 | MDQ6VXNlcjE1MDk4NTk4 | User | false |
huggingface/transformers | 2,843,001,264 | I_kwDOCUB6oc6pdMGw | 36,119 | https://github.com/huggingface/transformers/issues/36119 | https://api.github.com/repos/huggingface/transformers/issues/36119 | Adapting Whisper to the new loss_function attribute | @ArthurZucker @muellerzr Following up on #35838, #34191, #34198, #34283
I would like to help bring in Whisper into this. I see it was not included in the last #35875 round of fixes related to the loss function bug fix (grad acc.) nor the new global "loss_function" attr. Being an encodec model derived from Bart code in... | open | null | false | 11 | [
"Good Second Issue",
"Feature request"
] | [] | 2025-02-10T16:37:08Z | 2026-06-19T11:12:20Z | null | NONE | null | 20260619T120025Z | 2026-06-19T12:00:25Z | yoadsn | 1,684,709 | MDQ6VXNlcjE2ODQ3MDk= | User | false |
huggingface/transformers | 3,545,702,940 | I_kwDOCUB6oc7TVyIc | 41,826 | https://github.com/huggingface/transformers/issues/41826 | https://api.github.com/repos/huggingface/transformers/issues/41826 | Integrating TiledMLP for a much smaller memory footprint | Similar to `TiledFusedLogitsLoss` https://github.com/huggingface/transformers/issues/41306 it's of a great benefit to perform `TiledMLP` as well, which can save a lot of HBM memory and allow longer seqlen/larger batch size or even enable a bs=1 training which wasn't possible before.
This technology comes from the Arct... | open | null | false | 5 | [
"Feature request"
] | [] | 2025-10-23T17:07:21Z | 2026-06-17T08:33:40Z | null | CONTRIBUTOR | null | 20260617T120115Z | 2026-06-17T12:01:15Z | stas00 | 10,676,103 | MDQ6VXNlcjEwNjc2MTAz | User | false |
huggingface/transformers | 4,681,044,495 | I_kwDOCUB6oc8AAAABFwMSDw | 46,711 | https://github.com/huggingface/transformers/issues/46711 | https://api.github.com/repos/huggingface/transformers/issues/46711 | Clarify Gemma4 vision bidirectional mask behavior for full vs sliding attention | ## Description
I noticed a potential mismatch between HuggingFace Transformers' Gemma4 vision bidirectional mask behavior and the official [`google-deepmind/gemma`](https://github.com/google-deepmind/gemma/tree/main) implementation.
In the official `google/gemma` implementation, `use_bidirectional_attention="vision"`... | closed | completed | false | 2 | [] | [] | 2026-06-17T07:21:17Z | 2026-06-30T09:56:24Z | 2026-06-30T09:56:24Z | NONE | null | 20260630T120021Z | 2026-06-30T12:00:21Z | wnma3mz | 23,001,152 | MDQ6VXNlcjIzMDAxMTUy | User | false |
huggingface/transformers | 4,680,961,599 | I_kwDOCUB6oc8AAAABFwHOPw | 46,710 | https://github.com/huggingface/transformers/issues/46710 | https://api.github.com/repos/huggingface/transformers/issues/46710 | Accuracy regression in DeepSeek-R1-Distill-Llama-8B after upgrading Transformers from 4.55 to 5.9 | ### System Info
transformers: 4.55 vs 5.9.0
PyTorch: 2.13
### Who can help?
@ArthurZucker and @itazap
### 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... | closed | completed | false | 3 | [
"bug"
] | [] | 2026-06-17T07:08:17Z | 2026-06-23T02:26:33Z | 2026-06-23T02:26:33Z | NONE | null | 20260623T060018Z | 2026-06-23T06:00:18Z | xiaowangintel | 109,140,002 | U_kgDOBoFYIg | User | false |
huggingface/transformers | 2,829,863,352 | I_kwDOCUB6oc6orEm4 | 36,028 | https://github.com/huggingface/transformers/issues/36028 | https://api.github.com/repos/huggingface/transformers/issues/36028 | [Beam Search Optimization] Memory-Efficient Beam Search for multi-head attention | ### Feature request
Adding a [Trie-Based Beam Search](https://www.arxiv.org/abs/2502.00085) implementation.
### Motivation
The current beam search implementation in Hugging Face's transformers library maintains separate KV caches for each beam candidate, even when they share common prefixes. For example, with a bea... | open | null | false | 9 | [
"Feature request",
"Generation"
] | [] | 2025-02-04T11:14:11Z | 2026-06-22T01:54:08Z | null | NONE | null | 20260622T060015Z | 2026-06-22T06:00:15Z | brian030128 | 37,238,439 | MDQ6VXNlcjM3MjM4NDM5 | User | false |
huggingface/transformers | 4,684,387,871 | I_kwDOCUB6oc8AAAABFzYWHw | 46,723 | https://github.com/huggingface/transformers/issues/46723 | https://api.github.com/repos/huggingface/transformers/issues/46723 | Several models hardcode `torch.float64` in their forward pass, crashing on MPS (Apple Silicon) | ### Summary
While auditing recent models for Apple Silicon (MPS) compatibility, I found two models that crash during a normal forward pass on MPS because they hardcode `torch.float64`, which the MPS backend does not support (`TypeError: Cannot convert a MPS Tensor to float64 dtype as the MPS framework doesn't support ... | open | null | false | 1 | [] | [] | 2026-06-17T14:51:13Z | 2026-06-17T15:46:35Z | null | CONTRIBUTOR | null | 20260617T180027Z | 2026-06-17T18:00:27Z | qflen | 194,738,340 | U_kgDOC5t4pA | User | false |
huggingface/transformers | 4,684,892,701 | I_kwDOCUB6oc8AAAABFz3KHQ | 46,726 | https://github.com/huggingface/transformers/issues/46726 | https://api.github.com/repos/huggingface/transformers/issues/46726 | AutoProcessor reports “Unrecognized image processor” when required vision backend is missing | ### System Info
Observed with `transformers==5.12.1` in an environment with `torch` installed but `torchvision` missing.
### Who can help?
_No response_
### Information
- [ ] The official example scripts
- [x] My own modified scripts
### Tasks
- [x] An officially supported task in the `examples` folder (such as ... | closed | completed | false | 1 | [
"bug"
] | [] | 2026-06-17T15:54:08Z | 2026-06-18T11:31:02Z | 2026-06-18T11:31:02Z | CONTRIBUTOR | null | 20260618T120025Z | 2026-06-18T12:00:25Z | sisaman | 3,238,345 | MDQ6VXNlcjMyMzgzNDU= | User | false |
huggingface/transformers | 4,685,032,799 | I_kwDOCUB6oc8AAAABFz_tXw | 46,728 | https://github.com/huggingface/transformers/issues/46728 | https://api.github.com/repos/huggingface/transformers/issues/46728 | Aria: get_number_of_image_patches returns wrong count with max_image_size=980 when best resolution has odd 490 multiples | ## System Info
```
transformers version: main (latest)
Python: 3.10
torch: 2.6
```
## Who can help?
@zucchini-nlp @yonigozlan (Aria model maintainers)
## Information
- [x] The official example scripts
- [ ] My own modified scripts
## Tasks
- [ ] An officially supported task in the `examples` folder
- [x] My own ... | closed | completed | false | 1 | [] | [] | 2026-06-17T16:12:26Z | 2026-06-19T10:10:37Z | 2026-06-19T10:10:37Z | CONTRIBUTOR | null | 20260619T120025Z | 2026-06-19T12:00:25Z | arnavkewalram | 154,707,184 | U_kgDOCTik8A | User | false |
huggingface/transformers | 4,685,629,717 | I_kwDOCUB6oc8AAAABF0kJFQ | 46,731 | https://github.com/huggingface/transformers/issues/46731 | https://api.github.com/repos/huggingface/transformers/issues/46731 | LabelSmoother ignores num_items_in_batch, inflating gradients by ~gradient_accumulation_steps with label smoothing | ### System Info
- `transformers` 5.13.0.dev0 (main, 9fd7b6789d; also reproduced on 36193bf)
- `torch` 2.11.0, Python 3.13.5, macOS 15.6 arm64 (forced CPU via `use_cpu=True`; not device-specific — also reproduces on the default MPS device)
### Who can help?
@muellerzr @SunMarc
### Reproduction
```python
import torc... | open | null | false | 0 | [] | [] | 2026-06-17T17:35:31Z | 2026-06-17T17:35:31Z | null | CONTRIBUTOR | null | 20260617T180027Z | 2026-06-17T18:00:27Z | Incheonkirin | 42,427,560 | MDQ6VXNlcjQyNDI3NTYw | User | false |
huggingface/transformers | 4,686,485,878 | I_kwDOCUB6oc8AAAABF1YZdg | 46,736 | https://github.com/huggingface/transformers/issues/46736 | https://api.github.com/repos/huggingface/transformers/issues/46736 | LoRA fine-tuning of an FP8 checkpoint is blocked: `get_peft_model` does not clear `validate_quantization_for_training` (QuantizationMethod.FP8) | ### System Info
- `transformers` 5.9.0
- `peft` 0.19.1
- `accelerate` 1.13.0
- `torch` 2.9.1 (CUDA 12.9)
- Python 3.12
- Platform: SageMaker training job, 2× p5.48xlarge (16× H100 80GB), FSDP `full_shard`, `attn_implementation="eager"`
- Model: `deepseek-ai/DeepSeek-V4-Flash` (284B MoE, published **FP8-only** — `quant... | open | null | false | 4 | [] | [] | 2026-06-17T19:38:43Z | 2026-06-24T14:11:05Z | null | NONE | null | 20260624T180019Z | 2026-06-24T18:00:19Z | brunopistone | 10,196,125 | MDQ6VXNlcjEwMTk2MTI1 | User | false |
huggingface/transformers | 4,690,378,812 | I_kwDOCUB6oc8AAAABF5GAPA | 46,739 | https://github.com/huggingface/transformers/issues/46739 | https://api.github.com/repos/huggingface/transformers/issues/46739 | GlmMoeDsa: head_dim→qk_rope_head_dim attribute_map silently clobbers the rope head dim (GLM-5.2) | ### System Info
- `transformers` 5.8.1
- Model: GLM-5.2 (`model_type: glm_moe_dsa`, `GlmMoeDsaForCausalLM`)
### Description
`GlmMoeDsaConfig` declares `attribute_map = {"head_dim": "qk_rope_head_dim"}`, so `head_dim` is an
alias for `qk_rope_head_dim`. But GLM-5.2's released `config.json` sets **both** fields with
*... | closed | completed | false | 2 | [] | [] | 2026-06-18T07:56:59Z | 2026-06-18T08:33:49Z | 2026-06-18T08:33:49Z | NONE | null | 20260618T120025Z | 2026-06-18T12:00:25Z | Kasempiternal | 36,620,658 | MDQ6VXNlcjM2NjIwNjU4 | User | false |
huggingface/transformers | 4,695,139,012 | I_kwDOCUB6oc8AAAABF9oixA | 46,752 | https://github.com/huggingface/transformers/issues/46752 | https://api.github.com/repos/huggingface/transformers/issues/46752 | `apply_chat_template` crashes with `IndexError` when passed an empty list | ### System Info
transformers: 5.13.0.dev0 (latest main, commit 2b422bde64)
Python: 3.14.4
OS: Ubuntu 25.04 (WSL2)
### 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/SQ... | closed | completed | false | 0 | [
"bug"
] | [] | 2026-06-18T18:50:23Z | 2026-06-22T12:17:53Z | 2026-06-22T12:17:53Z | NONE | null | 20260622T180028Z | 2026-06-22T18:00:28Z | sharmax-vikas | 175,734,208 | U_kgDOCnl9wA | User | false |
huggingface/transformers | 4,695,715,743 | I_kwDOCUB6oc8AAAABF-Lvnw | 46,756 | https://github.com/huggingface/transformers/issues/46756 | https://api.github.com/repos/huggingface/transformers/issues/46756 | FR: Implement Downstream Integration Smoke Tests in CI/CD to Prevent Ecosystem Regression | ### Feature request
We request the addition of Downstream Integration Smoke Tests within the transformers CI/CD automation pipeline (GitHub Actions/CircleCI). Before a new release is cut, the test matrix should pull down and execute basic smoke generation scripts against a select group of foundational, high-impact d... | closed | completed | false | 4 | [
"Feature request"
] | [] | 2026-06-18T20:22:06Z | 2026-06-20T03:57:56Z | 2026-06-19T13:20:01Z | NONE | null | 20260620T060014Z | 2026-06-20T06:00:14Z | briancullinan2 | 112,788,814 | U_kgDOBrkFTg | User | false |
huggingface/transformers | 4,697,723,189 | I_kwDOCUB6oc8AAAABGAGRNQ | 46,762 | https://github.com/huggingface/transformers/issues/46762 | https://api.github.com/repos/huggingface/transformers/issues/46762 | Substantial numerical differences with Minimax MSA attention vs Minimax's Reference | ### System Info
-> Minimax-M3/references/transformers/modular_minimax_m3_vl.py:550
The problematic operation is:
```
block_scores = scores.amax(dim=-1).amax(dim=1)
```
That first max pools within the 128-token block, which is expected. The second max, amax(dim=1), collapses H_idx, producing one shared selected b... | closed | completed | false | 2 | [
"bug"
] | [] | 2026-06-19T03:38:05Z | 2026-06-20T19:54:43Z | 2026-06-20T19:54:43Z | NONE | null | 20260621T000031Z | 2026-06-21T00:00:31Z | CoffeeVampir3 | 48,565,901 | MDQ6VXNlcjQ4NTY1OTAx | User | false |
huggingface/transformers | 2,207,518,660 | I_kwDOCUB6oc6DlA_E | 29,870 | https://github.com/huggingface/transformers/issues/29870 | https://api.github.com/repos/huggingface/transformers/issues/29870 | Bart evaluation throws the following error at generate(): UnboundLocalError: 'model_kwargs['decoder_attention_mask']' is used before assignment | ### System Info
- `transformers` version: 4.39.0
- Platform: Linux-5.4.0-167-generic-x86_64-with-glibc2.31
- Python version: 3.10.13
- Huggingface_hub version: 0.21.4
- Safetensors version: 0.4.2
- Accelerate version: 0.28.0
- Accelerate config: not found
- PyTorch version (GPU?): 2.2.1+cu121 (False)
- Te... | closed | completed | false | 5 | [
"TensorFlow",
"Examples",
"Good Second Issue"
] | [] | 2024-03-26T08:02:38Z | 2026-06-19T13:14:56Z | 2026-06-19T13:14:56Z | NONE | null | 20260619T180015Z | 2026-06-19T18:00:15Z | Madhumitha-MCW | 109,727,353 | U_kgDOBopOeQ | User | false |
huggingface/transformers | 4,701,883,422 | I_kwDOCUB6oc8AAAABGEEMHg | 46,772 | https://github.com/huggingface/transformers/issues/46772 | https://api.github.com/repos/huggingface/transformers/issues/46772 | Nvidia ModelOpt (NVFP4) compatibility for DiffusionGemma | ### Feature request
Currently, when I try to run DiffusionGemma in NVFP4 through transformers I see the warning
```
[transformers] Unknown quantization type, got modelopt - supported types are: ['awq', ..., 'gemma']
```
After installing `nvidia-modelopt[hf]` transformers is downgraded.
- nvidia-modelopt[hf] instal... | closed | not_planned | false | 1 | [
"Feature request"
] | [] | 2026-06-19T15:33:18Z | 2026-06-19T16:12:43Z | 2026-06-19T16:12:00Z | NONE | null | 20260619T180015Z | 2026-06-19T18:00:15Z | ppabis | 121,432,819 | U_kgDOBzzq8w | User | false |
huggingface/transformers | 3,290,275,644 | I_kwDOCUB6oc7EHZ88 | 39,893 | https://github.com/huggingface/transformers/issues/39893 | https://api.github.com/repos/huggingface/transformers/issues/39893 | Add VideoPrism | ### Model description
I'd like to contribute the VideoPrism model by google deepmind. The code base (written in flax) was made public in June and the base model weights alone has 2.7k downloads on [HF hub](https://huggingface.co/google/videoprism-base-f16r288) with almost no support except for the original repo. There... | closed | completed | false | 3 | [
"New model"
] | [] | 2025-08-04T17:35:42Z | 2026-06-19T18:03:03Z | 2026-06-19T18:03:03Z | CONTRIBUTOR | null | 20260620T000023Z | 2026-06-20T00:00:23Z | MHRDYN7 | 113,298,714 | U_kgDOBsDNGg | User | false |
huggingface/transformers | 4,707,829,708 | I_kwDOCUB6oc8AAAABGJvHzA | 46,785 | https://github.com/huggingface/transformers/issues/46785 | https://api.github.com/repos/huggingface/transformers/issues/46785 | Is transformers 5.12.1 requires tokenizers<=0.23.0,>=0.22.0 really right? | https://github.com/huggingface/tokenizers/releases/tag/v0.23.1 says that:
tokenizers 0.23.1 is the first proper stable release in the 0.23 line — 0.23.0 only ever shipped as rc0 because the release pipeline itself was broken (Node side hadn't shipped multi-platform binaries since 2023, Python side was on pyo3 0.27 wit... | open | null | false | 4 | [] | [] | 2026-06-20T17:59:46Z | 2026-06-29T09:01:40Z | null | NONE | null | 20260629T120018Z | 2026-06-29T12:00:18Z | jrp2014 | 8,142,876 | MDQ6VXNlcjgxNDI4NzY= | User | false |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.