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 | 1,050,733,132 | I_kwDOCUB6oc4-oOpM | 14,368 | https://github.com/huggingface/transformers/issues/14368 | https://api.github.com/repos/huggingface/transformers/issues/14368 | Export LayoutLMv2 to onnx | I am trying to export LayoutLMv2 model to onnx but there is no support for that available in transformers library.
I have tried to follow the method available for layoutLM but that is not working.
Here is config class for LayoutLMv2
```
class LayoutLMv2OnnxConfig(OnnxConfig):
def __init__(
self,
config: PretrainedConfig,
task: str = "default",
patching_specs: List[PatchingSpec] = None,
):
super().__init__(config, task=task, patching_specs=patching_specs)
self.max_2d_positions = config.max_2d_position_embeddings - 1
@property
def inputs(self) -> Mapping[str, Mapping[int, str]]:
return OrderedDict(
[
("input_ids", {0: "batch", 1: "sequence"}),
("bbox", {0: "batch", 1: "sequence"}),
("image", {0: "batch", 1: "sequence"}),
("attention_mask", {0: "batch", 1: "sequence"}),
("token_type_ids", {0: "batch", 1: "sequence"}),
]
)
def generate_dummy_inputs(
self,
tokenizer: PreTrainedTokenizer,
batch_size: int = -1,
seq_length: int = -1,
is_pair: bool = False,
framework: Optional[TensorType] = None,
) -> Mapping[str, Any]:
"""
Generate inputs to provide to the ONNX exporter for the specific framework
Args:
tokenizer: The tokenizer associated with this model configuration
batch_size: The batch size (int) to export the model for (-1 means dynamic axis)
seq_length: The sequence length (int) to export the model for (-1 means dynamic axis)
is_pair: Indicate if the input is a pair (sentence 1, sentence 2)
framework: The framework (optional) the tokenizer will generate tensor for
Returns:
Mapping[str, Tensor] holding the kwargs to provide to the model's forward function
"""
input_dict = super().generate_dummy_inputs(tokenizer, batch_size, seq_length, is_pair, framework)
# Generate a dummy bbox
box = [48, 84, 73, 128]
if not framework == TensorType.PYTORCH:
raise NotImplementedError("Exporting LayoutLM to ONNX is currently only supported for PyTorch.")
if not is_torch_available():
raise ValueError("Cannot generate dummy inputs without PyTorch installed.")
import torch
batch_size, seq_length = input_dict["input_ids"].shape
input_dict["bbox"] = torch.tensor([*[box] * seq_length]).tile(batch_size, 1, 1)
return input_dict
onnx_config = LayoutLMv2OnnxConfig(model.config)
export(tokenizer=tokenizer, model=model, config=onnx_config, opset=12, output=Path('onnx/layoutlmv2.onnx'))
```
Running the export line is raising this error,
```
---------------------------------------------------------------------------
AssertionError Traceback (most recent call last)
<ipython-input-25-99a1f167e396> in <module>()
----> 1 export(tokenizer=tokenizer, model=model, config=onnx_config, opset=12, output=Path('onnx/layoutlmv2.onnx'))
3 frames
/usr/local/lib/python3.7/dist-packages/transformers/models/layoutlmv2/tokenization_layoutlmv2.py in __call__(self, text, text_pair, boxes, word_labels, add_special_tokens, padding, truncation, max_length, stride, pad_to_multiple_of, return_tensors, return_token_type_ids, return_attention_mask, return_overflowing_tokens, return_special_tokens_mask, return_offsets_mapping, return_length, verbose, **kwargs)
449
450 words = text if text_pair is None else text_pair
--> 451 assert boxes is not None, "You must provide corresponding bounding boxes"
452 if is_batched:
453 assert len(words) == len(boxes), "You must provide words and boxes for an equal amount of examples"
AssertionError: You must provide corresponding bounding boxes
``` | closed | completed | false | 28 | [
"Good First Issue"
] | [] | 2021-11-11T08:54:39Z | 2026-03-20T08:32:38Z | 2026-03-20T08:32:16Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | fadi212 | 37,739,280 | MDQ6VXNlcjM3NzM5Mjgw | User | false |
huggingface/transformers | 2,363,874,975 | I_kwDOCUB6oc6M5d6f | 31,515 | https://github.com/huggingface/transformers/issues/31515 | https://api.github.com/repos/huggingface/transformers/issues/31515 | from_pretrained 加载checkpoint过慢的问题 | ### System Info
latest python3.9.8
### 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, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
我用ollama加载模型明显比上述方法快，有没有办法可以提高加载速度
### Expected behavior
1 | closed | completed | false | 3 | [] | [] | 2024-06-20T08:41:06Z | 2026-03-20T03:44:30Z | 2024-07-29T08:04:21Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | zhaoyuchen1128 | 167,266,669 | U_kgDOCfhJbQ | User | false |
huggingface/transformers | 2,418,835,728 | I_kwDOCUB6oc6QLIEQ | 32,090 | https://github.com/huggingface/transformers/issues/32090 | https://api.github.com/repos/huggingface/transformers/issues/32090 | [Error] with Trainer: TypeError: Unsupported types (<class 'NoneType'>) passed to `_gpu_broadcast_one`. | ### System Info
- `transformers` version: 4.42.4
- Platform: Linux-5.15.0-101-generic-x86_64-with-glibc2.35
- Python version: 3.10.13
- Huggingface_hub version: 0.24.0
- Safetensors version: 0.4.2
- Accelerate version: 0.32.0
- Accelerate config: not found
- PyTorch version (GPU?): 2.3.1+cu121 (False)
- Tensorflow version (GPU?): not installed (NA)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Using distributed or parallel set-up in script?: <fill in>
### Who can help?
@muellerzr @SunMarc @ArthurZucker
### 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)
### Reproduction
https://gist.github.com/halixness/eadd6d1d89ae48597f70cb09f2b44139
### Expected behavior
Hello,
I have written a simple training script to train from scratch a gpt2-like model with a large dataset of strings (molecules in SMILES format). After around ~2k steps (`batch_size=128`, `#samples = ~1.5M`), I encounter the following error:
```
TypeError: Unsupported types (<class 'NoneType'>) passed to `_gpu_broadcast_one`. Only nested list/tuple/dicts of objects that are valid for `is_torch_tensor` should be passed.
```
I tried already:
- to use the `default_data_collator` instead and to manually group samples as in [in the official example](https://github.com/huggingface/transformers/blob/89575b567e061fd87bdd655ba188b6c7a922d54a/examples/pytorch/language-modeling/run_clm.py#L513).
- to check manually for the value of the batch that makes the script crash apparently: no NaN values, it all seems to make sense.
- to check whether the dataset initially contains any empty or None strings, which is not the case.
I'm not sure about what could case this error. Any suggestion is much appreciated! | closed | completed | false | 3 | [
"trainer",
"bug"
] | [] | 2024-07-19T13:01:37Z | 2026-03-20T06:22:10Z | 2024-09-22T08:06:59Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | halixness | 20,798,848 | MDQ6VXNlcjIwNzk4ODQ4 | User | false |
huggingface/transformers | 2,914,781,972 | I_kwDOCUB6oc6tvAsU | 36,683 | https://github.com/huggingface/transformers/issues/36683 | https://api.github.com/repos/huggingface/transformers/issues/36683 | AttributeError: 'Gemma3Config' object has no attribute 'vocab_size' | ### System Info
v4.50.0.dev0
### Who can help?
@ArthurZucker
@LysandreJik
@xenova
### 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)
### Reproduction
I am trying to run the new Gemma3 model, using version '4.50.0.dev0'.
When loading the model I get the error: 'Gemma3Config' object has no attribute 'vocab_size'.
Looking into this it seems Gemma3Config has `vocab_size` nested in a "text_config" attribute.
I try to load the model as AutoModelForCausalLM, running it with Gemma3ForConditionalGeneration does not raise this issue. Am I wrong in assuming I can run Gemma 3 as AutoModelForCausalLM?
### Expected behavior
Loading the model as AutoModelForCausalLM.from_pretrained without issue. | open | reopened | false | 39 | [
"bug"
] | [] | 2025-03-12T18:11:39Z | 2026-03-23T13:36:49Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | jumelet | 9,407,977 | MDQ6VXNlcjk0MDc5Nzc= | User | false |
huggingface/transformers | 3,229,815,847 | I_kwDOCUB6oc7AgxQn | 39,404 | https://github.com/huggingface/transformers/issues/39404 | https://api.github.com/repos/huggingface/transformers/issues/39404 | Whisper `return_language` with pipeline no longer working | ### System Info
Platform: Initially discovered on Nvidia. Can be reproduced on CPU and in Google Colab (see attached gist).
- `transformers` version: 4.53.2
- Platform: Linux-6.6.87.2-microsoft-standard-WSL2-x86_64-with-glibc2.39
- Python version: 3.12.3
- Huggingface_hub version: 0.33.4
- Safetensors version: 0.5.3
- Accelerate version: 1.8.1
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.7.1+cu126 (CUDA)
- Tensorflow version (GPU?): not installed (NA)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Using distributed or parallel set-up in script?: No
- Using GPU in script?: Yes and No.
- GPU type: NVIDIA GeForce RTX 3090
### Who can help?
@eustlb @ArthurZucker
### 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)
### Reproduction
<s>Sometime between `transformers==4.46.3` and `transfomers==4.53.2 (latest as of now)`,</s> At #34135, the `return_language` argument for pipeline stopped working. The ending timestamp for the last word is also missing.
Example (exported from Google Colab): https://gist.github.com/Metric-Void/ce2b9fe2faed0cdf6e5fd328599fd4c7
Code for testing:
```
import torch
from transformers import pipeline
from transformers.configuration_utils import PretrainedConfig
pipeline = pipeline(
task="automatic-speech-recognition",
model="openai/whisper-tiny",
torch_dtype=torch.float16,
config=PretrainedConfig(
attn_implementation="flash_attention_2"
)
)
result = pipeline("https://huggingface.co/datasets/Narsil/asr_dummy/resolve/main/mlk.flac", return_language=True, return_timestamps='word')
result["chunks"]
```
Before (`transformers==4.46.3`):
```
[{'text': ' I', 'timestamp': (1.04, 1.36), 'language': 'english'},
{'text': ' have', 'timestamp': (1.36, 1.68), 'language': 'english'},
{'text': ' a', 'timestamp': (1.68, 1.94), 'language': 'english'},
{'text': ' dream.', 'timestamp': (1.94, 3.82), 'language': 'english'},
{'text': ' Good', 'timestamp': (3.82, 3.98), 'language': 'english'},
{'text': ' one', 'timestamp': (3.98, 4.16), 'language': 'english'},
{'text': ' day.', 'timestamp': (4.16, 6.4), 'language': 'english'},
{'text': ' This', 'timestamp': (6.4, 6.58), 'language': 'english'},
{'text': ' nation', 'timestamp': (6.58, 7.24), 'language': 'english'},
{'text': ' will', 'timestamp': (7.24, 7.82), 'language': 'english'},
{'text': ' rise', 'timestamp': (7.82, 8.3), 'language': 'english'},
{'text': ' up.', 'timestamp': (8.3, 10.3), 'language': 'english'},
{'text': ' Live', 'timestamp': (10.3, 10.56), 'language': 'english'},
{'text': ' out', 'timestamp': (10.56, 10.98), 'language': 'english'},
{'text': ' the', 'timestamp': (10.98, 11.02), 'language': 'english'},
{'text': ' true', 'timestamp': (11.02, 11.3), 'language': 'english'},
{'text': ' meaning', 'timestamp': (11.3, 11.6), 'language': 'english'},
{'text': ' of', 'timestamp': (11.6, 11.86), 'language': 'english'},
{'text': ' its', 'timestamp': (11.86, 12.08), 'language': 'english'},
{'text': ' dream.', 'timestamp': (12.08, 12.98), 'language': 'english'}]
```
After (`transfomers==4.53.2`):
```
[{'text': ' I', 'timestamp': (1.04, 1.36), 'language': None},
{'text': ' have', 'timestamp': (1.36, 1.68), 'language': None},
{'text': ' a', 'timestamp': (1.68, 1.94), 'language': None},
{'text': ' dream.', 'timestamp': (1.94, 3.82), 'language': None},
{'text': ' But', 'timestamp': (3.82, 3.96), 'language': None},
{'text': ' one', 'timestamp': (3.96, 4.18), 'language': None},
{'text': ' day,', 'timestamp': (4.18, 6.22), 'language': None},
{'text': ' this', 'timestamp': (6.22, 6.58), 'language': None},
{'text': ' nation', 'timestamp': (6.58, 7.22), 'language': None},
{'text': ' will', 'timestamp': (7.22, 7.82), 'language': None},
{'text': ' rise', 'timestamp': (7.82, 8.3), 'language': None},
{'text': ' up,', 'timestamp': (8.3, 10.2), 'language': None},
{'text': ' live', 'timestamp': (10.2, 10.56), 'language': None},
{'text': ' out', 'timestamp': (10.56, 10.98), 'language': None},
{'text': ' the', 'timestamp': (10.98, 11.02), 'language': None},
{'text': ' true', 'timestamp': (11.02, 11.3), 'language': None},
{'text': ' meaning', 'timestamp': (11.3, 11.6), 'language': None},
{'text': ' of', 'timestamp': (11.6, 11.86), 'language': None},
{'text': ' its', 'timestamp': (11.86, 12.08), 'language': None},
{'text': ' dream.', 'timestamp': (12.08, None), 'language': None}]
```
### Expected behavior
The old behaviour was correct.
Maybe related: #21311, #21427, #25138, #27604, #29520, #31572 | open | reopened | false | 12 | [
"bug",
"Audio"
] | [
"eustlb"
] | 2025-07-14T19:36:46Z | 2026-03-24T13:00:45Z | null | NONE | null | 20260324T143453Z | 2026-03-24T14:34:53Z | Metric-Void | 21,335,640 | MDQ6VXNlcjIxMzM1NjQw | User | false |
huggingface/transformers | 3,722,400,306 | I_kwDOCUB6oc7d31Iy | 42,832 | https://github.com/huggingface/transformers/issues/42832 | https://api.github.com/repos/huggingface/transformers/issues/42832 | Question about tie_weights | Hi,
I noticed that the logic of the tie_weights function has changed in the transformers 5.0.0rc.
In v4.x, when tie_word_embeddings=True, weights between embed_tokens.weight and lm_head.weight were always tied, regardless of whether both tensors were present in the checkpoint.
However, in v5.0.0rc, if both embed_tokens.weight and lm_head.weight are explicitly present in the checkpoint, the model no longer ties them, resulting in two independent copies of the weights.
https://github.com/huggingface/transformers/blob/v5.0.0rc1/src/transformers/modeling_utils.py#L2362-L2369
This change affects training behavior: gradients will now be computed and applied separately to the two weight matrices, potentially leading to divergent updates and inconsistent model behavior compared to previous versions.
I'm concerned this might introduce subtle training instabilities or break backward compatibility for users who expect tied embeddings and output layers by default (as was the case in v4.x). I’m wondering whether this change could lead to any unexpected problems？For example, although embed_tokens.weight and lm_head. weight exist in the checkpoint, I still want to tie these two weights.
Thanks! | open | null | false | 12 | [] | [] | 2025-12-12T07:52:43Z | 2026-03-24T08:12:04Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | cjw-d | 73,046,570 | MDQ6VXNlcjczMDQ2NTcw | User | false |
huggingface/transformers | 3,843,302,597 | I_kwDOCUB6oc7lFCTF | 43,412 | https://github.com/huggingface/transformers/issues/43412 | https://api.github.com/repos/huggingface/transformers/issues/43412 | gemma3n executorch export fails missing self training guard and erfinv not supported | system info
- transformers 5.0.0rc1
- torch 2.9.1
- executorch 1.0.1
- python 3.11
- ubuntu 24.04
who can help
@ArthurZucker @younesbelkada
information
- the official example scripts
tasks
- an officially supported task in the examples folder
problem description
when exporting gemma3n models to executorch pte format using torch export there are two bugs that prevent successful export
issue 1 missing self training guard causes constant is mutated error
file transformers/models/gemma3n/modeling_gemma3n.py line 1119
the altup_coef_clip clamp operation runs during inference causing torch export to fail with cannot mutate a constant error
current code
```python
if self.config.altup_coef_clip is not None:
self.correction_coefs.weight.data.clamp_(...)
```
fix
```python
if self.training and self.config.altup_coef_clip is not None:
self.correction_coefs.weight.data.clamp_(...)
```
issue 2 dynamic erfinv not supported by executorch
file transformers/models/gemma3n/modeling_gemma3n.py gemma3ntextmlp class
the _gaussian_topk method calls torch distributions normal Normal icdf during forward pass which uses erfinv an operator not in the executorch core aten opset
fix pre compute std_multiplier at init time instead of during forward
```python
# in __init__ add after self.activation_sparsity = ...
if self.activation_sparsity > 0.0:
normal_dist = torch.distributions.normal.Normal(0, 1)
std_multiplier = normal_dist.icdf(torch.tensor(self.activation_sparsity, dtype=torch.float32))
self.register_buffer("_std_multiplier", std_multiplier, persistent=False)
# replace _gaussian_topk method
def _gaussian_topk(self, inputs: torch.Tensor) -> torch.Tensor:
std_multiplier = self._std_multiplier.to(inputs.dtype)
inputs_mean = torch.mean(inputs, dim=-1, keepdim=True)
inputs_std = torch.std(inputs, dim=-1, keepdim=True, unbiased=False)
cutoff_x = inputs_mean + inputs_std * std_multiplier
return nn.functional.relu(inputs - cutoff_x)
```
expected behavior
gemma3n models should export to executorch pte format without requiring manual patches
reproduction
```bash
pip install torch transformers executorch
```
```python
from transformers import AutoModelForCausalLM
import torch
from executorch.exir import to_edge, EdgeCompileConfig
model = AutoModelForCausalLM.from_pretrained('google/gemma-3n-E2B-it', torch_dtype=torch.float16)
text_model = model.language_model
text_model.lm_head = model.lm_head
ep = torch.export.export(text_model, (torch.zeros(1, 32, dtype=torch.long),), strict=False)
et = to_edge(ep, compile_config=EdgeCompileConfig(_check_ir_validity=False))
et.to_executorch()
```
additional context
successfully exported both e2b 8.4gb and e4b 13.1gb text only models after applying these patches
reproduction repo https://github.com/maceip/gemma3n-executorch
exported models
- https://huggingface.co/macmacmacmac/gemma-3n-E2B-it-pte
- https://huggingface.co/macmacmacmac/gemma-3n-E4B-it-pte
- https://huggingface.co/macmacmacmac/odml-bench | open | null | false | 5 | [] | [] | 2026-01-22T14:51:37Z | 2026-03-20T08:08:47Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | maceip | 804,368 | MDQ6VXNlcjgwNDM2OA== | User | false |
huggingface/transformers | 3,871,334,543 | I_kwDOCUB6oc7mv-CP | 43,604 | https://github.com/huggingface/transformers/issues/43604 | https://api.github.com/repos/huggingface/transformers/issues/43604 | Revisit the condition for scaling the loss | Gradient accumulation requires scaled loss. Normally, loss scaling in the `Trainer` class depends on whether the
model accepts loss-related kwargs.
https://github.com/huggingface/transformers/blob/e7a2c0cc3471df9df0dd3ee739d1e1e034d549e0/src/transformers/trainer.py#L3827-L3930
In most custom trainers, we compute our own loss, but we still want to enable loss scaling in the parent class. The only way to do it a bit fragile:
```python
# Gradient accumulation requires scaled loss. Normally, loss scaling in the parent class depends on whether the
# model accepts loss-related kwargs. Since we compute our own loss, this check is irrelevant. We set
# self.model_accepts_loss_kwargs to False to enable scaling.
self.model_accepts_loss_kwargs = False
```
check the TRL repo, there are many occurence like this:
https://github.com/huggingface/trl/blob/7a530ba6d2ebf4101aaeb8002cd51c9bde4fb721/trl/experimental/kto/kto_trainer.py#L687
also related https://github.com/huggingface/trl/issues/2617 | closed | completed | false | 1 | [] | [] | 2026-01-29T15:42:54Z | 2026-03-20T08:08:40Z | 2026-03-20T08:08:40Z | MEMBER | null | 20260324T120231Z | 2026-03-24T12:02:31Z | qgallouedec | 45,557,362 | MDQ6VXNlcjQ1NTU3MzYy | User | false |
huggingface/transformers | 3,903,764,009 | I_kwDOCUB6oc7orrYp | 43,782 | https://github.com/huggingface/transformers/issues/43782 | https://api.github.com/repos/huggingface/transformers/issues/43782 | Qwen3VLForConditionalGeneration.from_pretrained weight_only = True error | ### System Info
platform: ubuntu 24.04.03
python 3.10
transformers 5.0
torch 2.10
accelerate 1.12
docker image: FROM nvidia/cuda:12.2.2-cudnn8-runtime-ubuntu22.04
https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct
### Who can help?
@yonigozlan @molbap
### Information
- [x] The official example scripts
- [ ] My own modified scripts
### Tasks
- [x] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
`
_pickle.UnpicklingError: Weights only load failed. In PyTorch 2.6, we changed the default value of the `weights_only` argument in `torch.load` from `False` to `True`. Re-running `torch.load` with `weights_only` set to `False` will likely succeed, but it can result in arbitrary code execution. Do it only if you got the file from a trusted source.
web-1 | Please file an issue with the following so that we can make `weights_only=True` compatible with your use case: WeightsUnpickler error:
`
I get the error when i try to run the code below
import logging
import os
import torch
from typing import Optional, Dict
from PIL import Image
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from app.core.config import settings
logger = logging.getLogger(__name__)
class LocalQwenVisionService:
_instance = None
def __new__(cls, model_path: Optional[str] = None):
if cls._instance is None:
cls._instance = super().__new__(cls)
cls._instance._initialized = False
return cls._instance
def __init__(self, model_path: Optional[str] = None):
if self._initialized:
return
self.model_path = model_path or settings.QWEN_MODEL_PATH
if not self.model_path or not os.path.exists(self.model_path):
raise ValueError(f"Model path invalid: {self.model_path}")
logger.info(f"🚀 Loading Qwen3-VL from: {self.model_path}")
self.model = Qwen3VLForConditionalGeneration.from_pretrained(
self.model_path,
dtype=torch.float16,
attn_implementation="sdpa",
local_files_only=True,
weights_only=False
)
self.processor = AutoProcessor.from_pretrained(
self.model_path,
local_files_only=True
)
logger.info(f"✅ Model loaded on: {next(self.model.parameters()).device}")
self._initialized = True
def analyze_image(
self,
image_path: str,
prompt: Optional[str] = None
) -> Dict:
"""✅ Analyze image and return response"""
if not os.path.exists(image_path):
return {"status": "error", "error": f"Image not found: {image_path}"}
if not prompt:
prompt = (
"Analyze this image and provide:\n"
"1. Description of content\n"
"2. Main objects and locations\n"
"3. Atmosphere and mood\n"
"4. Tags/categories\n"
"5. One-sentence summary"
)
try:
image = Image.open(image_path).convert("RGB")
messages = [
{
"role": "user",
"content": [
{"type": "image", "url": image},
{"type": "text", "text": prompt}
]
}
]
inputs = self.processor(
messages,
tokenize=True,
add_generation_prompt=True,
return_dict=True,
return_tensors='pt'
)
generated_ids = self.model.generate(
**inputs,
max_new_tokens=1024
)
generated_ids_trimmed = [
out_ids[len(in_ids):]
for in_ids, out_ids in zip(inputs.input_ids, generated_ids)
]
response_text = self.processor.batch_decode(
generated_ids_trimmed,
skip_special_tokens=True,
clean_up_tokenization_spaces=False
)[0]
return {
"status": "success",
"message": response_text
}
except Exception as e:
logger.error(f"❌ Error: {e}")
return {"status": "error", "error": str(e)}
def close(self):
"""Clean up resources"""
logger.info("🛑 Closing model")
if hasattr(self, "model"):
del self.model
torch.cuda.empty_cache()
What i tried
ENV TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 - to docker file didnt help
weights_only=False - inside Qwen3VLForConditionalGeneration.from_pretrained - didnt help
import numpy
import torch.serialization
torch.serialization.add_safe_globals([
(numpy._core.multiarray.scalar, 'numpy.core.multiarray.scalar'),
numpy.dtype,
numpy.dtypes.Float64DType
])
- didnt help
### Expected behavior
I have no clue where should i dig to fix the problem. | open | null | false | 3 | [
"bug"
] | [] | 2026-02-05T21:52:13Z | 2026-03-20T08:08:34Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | oscars17 | 57,970,391 | MDQ6VXNlcjU3OTcwMzkx | User | false |
huggingface/transformers | 3,911,668,331 | I_kwDOCUB6oc7pJ1Jr | 43,824 | https://github.com/huggingface/transformers/issues/43824 | https://api.github.com/repos/huggingface/transformers/issues/43824 | ImportError: cannot import name 'Qwen2_5_VLForConditionalGeneration' from 'transformers' | ### System Info
```
Traceback (most recent call last):
File "/home/wbrione/.conda/envs/watt_ai/bin/transformers", line 3, in <module>
from transformers.cli.transformers import main
File "/home/wbrione/.conda/envs/watt_ai/lib/python3.10/site-packages/transformers/cli/transformers.py", line 22, in <module>
from transformers.cli.serve import Serve
File "/home/wbrione/.conda/envs/watt_ai/lib/python3.10/site-packages/transformers/cli/serve.py", line 360, in <module>
class Serve:
File "/home/wbrione/.conda/envs/watt_ai/lib/python3.10/site-packages/transformers/cli/serve.py", line 588, in Serve
validator: TypeAdapter,
NameError: name 'TypeAdapter' is not defined
```
### Who can help?
@yonigozlan @molbap
### 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)
### Reproduction
```python
!pip install --upgrade --force-reinstall torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124
import os
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
import torch
import os
import json
import time
from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
# Load AWQ quantized model
# Note: Using "auto" for torch_dtype as recommended in the documentation
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
"Qwen/Qwen2.5-VL-72B-Instruct-AWQ",
torch_dtype = "auto",
device_map = "auto"
)
```
### Expected behavior
Import Qwen2 from transformers so I can run the rest of my code. | closed | completed | false | 11 | [
"bug"
] | [] | 2026-02-08T00:48:13Z | 2026-03-23T08:15:17Z | 2026-03-23T08:15:17Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | waltwalt36 | 197,971,906 | U_kgDOC8zPwg | User | false |
huggingface/transformers | 3,915,467,750 | I_kwDOCUB6oc7pYUvm | 43,856 | https://github.com/huggingface/transformers/issues/43856 | https://api.github.com/repos/huggingface/transformers/issues/43856 | Inefficient memory usage during Qwen3 MoE training | ### System Info
- `transformers` version: 4.57.3
- Platform: Linux-6.8.0-90-generic-x86_64-with-glibc2.35
- Python version: 3.12.12
- Huggingface_hub version: 0.36.0
- Safetensors version: 0.7.0
- Accelerate version: 1.12.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.9.0+cu126 (CUDA)
- Tensorflow version (GPU?): not installed (NA)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Using distributed or parallel set-up in script?: <fill in>
- Using GPU in script?: <fill in>
- GPU type: NVIDIA H100 80GB HBM3
### Who can help?
@ArthurZucker @Cyrilvallez
### Information
- [x] 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)
### Reproduction
[train_qlora_memory.py](https://github.com/user-attachments/files/25179495/train_qlora_memory.py)
[adapter_config.json](https://github.com/user-attachments/files/25179496/adapter_config.json)
Run with:
```
CUDA_VISIBLE_DEVICES=0 PYTORCH_ALLOC_CONF=expandable_segments:True,garbage_collection_threshold:0.8,max_split_size_mb:128 python train_qlora_memory.py "Qwen3/Qwen3-30B-A3B-Instruct-2507" --max_seq_length 1024 --batch_size 1 --max_steps 5 --path_config adapter_config.json --use_liger --use_gc
```
Output:
```
torch.OutOfMemoryError: CUDA out of memory. Tried to allocate 2.00 MiB. GPU 0 has a total capacity of 79.19 GiB of which 2.50 MiB is free. Process 2253701 has 79.18 GiB memory in use. Of the allocated memory 51.31 GiB is allocated by PyTorch, and 27.15 GiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation. See documentation for Memory Management (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)
```
Output confirming memory inefficiency: `27.15 GiB is reserved by PyTorch but unallocated`!
--
By the way, quantizing the model using **bitsandbytes** is also[ inefficient and leads to high memory consumption](https://github.com/bitsandbytes-foundation/bitsandbytes/issues/1849#issuecomment-3870385400): with **4-bit quantization** of the model **30B**, the actual memory consumption after loading the model is **46.75 GB** instead of the expected **16 GB**!
### Expected behavior
```
Loading checkpoint shards: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 16/16 [00:19<00:00, 1.24s/it]
LoraConfig(task_type=None, peft_type=<PeftType.LORA: 'LORA'>, auto_mapping=None, peft_version='0.18.0', base_model_name_or_path=None, revision=None, inference_mode=False, r=64, target_modules={'q_proj', 'o_proj', 'k_proj', 'v_proj', 'down_proj'}, exclude_modules=None, lora_alpha=64, lora_dropout=0.0, fan_in_fan_out=False, bias='none', use_rslora=False, modules_to_save=None, init_lora_weights=True, layers_to_transform=None, layers_pattern=None, rank_pattern={}, alpha_pattern={}, megatron_config=None, megatron_core='megatron.core', trainable_token_indices=None, loftq_config={}, eva_config=None, corda_config=None, use_dora=False, alora_invocation_tokens=None, use_qalora=False, qalora_group_size=16, layer_replication=None, runtime_config=LoraRuntimeConfig(ephemeral_gpu_offload=False), lora_bias=False, target_parameters=None, arrow_config=None, ensure_weight_tying=False)
trainable params: 1,160,773,632 || all params: 31,692,896,256 || trainable%: 3.6626
Gradient Checkpointing enabled
step 0 seq_length 1024 loss 11.360670 time 18.22s
step 1 seq_length 1024 loss 11.471191 time 14.65s
step 2 seq_length 1024 loss 11.253231 time 13.13s
step 3 seq_length 1024 loss 11.337120 time 11.62s
step 4 seq_length 1024 loss 11.379625 time 11.16s
cuda memory avg: 33269MB
cuda memory max: 45740MB
total time: 76.52s
file size: 4429.8MB
``` | closed | completed | false | 6 | [
"bug"
] | [] | 2026-02-09T09:48:03Z | 2026-03-23T08:15:13Z | 2026-03-23T08:15:13Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | yurkoff-mv | 82,467,993 | MDQ6VXNlcjgyNDY3OTkz | User | false |
huggingface/transformers | 3,925,350,527 | I_kwDOCUB6oc7p-Bh_ | 43,908 | https://github.com/huggingface/transformers/issues/43908 | https://api.github.com/repos/huggingface/transformers/issues/43908 | What is the process of adding a new hardware backend for Trainer? | I work in Qualcomm and we have a hardware backend like cuda / mps. I want to add it in the Trainer so that we can use Trainer class to perform training on our stack. What is the process of adding it?
We have a branch which currently adds the backend specific changes: https://github.com/quic-meetkuma/transformers/tree/qaic_support
Please guide on this. Will it be fine to directly raise a PR or there will be subsequent steps that I should be taking care of? | closed | completed | false | 5 | [] | [] | 2026-02-11T08:32:48Z | 2026-03-23T08:15:11Z | 2026-03-23T08:15:11Z | CONTRIBUTOR | null | 20260324T120231Z | 2026-03-24T12:02:31Z | quic-meetkuma | 200,747,495 | U_kgDOC_cp5w | User | false |
huggingface/transformers | 3,930,743,717 | I_kwDOCUB6oc7qSmOl | 43,939 | https://github.com/huggingface/transformers/issues/43939 | https://api.github.com/repos/huggingface/transformers/issues/43939 | Better regex in `build_glob_alternation` method | I notice in the following code:
https://github.com/huggingface/transformers/blob/4d5d49c34474be7cc2b6abd3179e7b317a17d8b1/src/transformers/core_model_loading.py#L67
https://github.com/huggingface/transformers/blob/4d5d49c34474be7cc2b6abd3179e7b317a17d8b1/src/transformers/core_model_loading.py#L75
The regex is formed by replace `*` with `.*`. I propose adding the escape character to signify the dot character `\.` and using the digit to better convey the message of looking for index in the model state dict keys `\d+`
The changes would be
```diff
-body = body.replace("*", r".*")
+body = body.replace(".", r"\.")
+body = body.replace("*", r"\d+")
```
I can add a PR to this if you think this would make more sense.
CC: @ArthurZucker
| closed | completed | false | 2 | [] | [] | 2026-02-12T09:21:30Z | 2026-03-23T08:15:08Z | 2026-03-23T08:15:08Z | MEMBER | null | 20260324T120231Z | 2026-03-24T12:02:31Z | ariG23498 | 36,856,589 | MDQ6VXNlcjM2ODU2NTg5 | User | false |
huggingface/transformers | 3,944,737,576 | I_kwDOCUB6oc7rH-so | 44,016 | https://github.com/huggingface/transformers/issues/44016 | https://api.github.com/repos/huggingface/transformers/issues/44016 | Syntax error in Transformer section 3 (Transformers, what can they do?) notebook | ### System Info
Build error due to syntax issue,
-> from transformers import pipeline
ner = pipeline("ner", grouped_entities=True)// remove parameter grouped_entities=True
ner("My name is Sylvain and I work at Hugging Face in Brooklyn.")
Error log: Notes:
- UNEXPECTED :can be ignored when loading from different task/architecture; not ok if you expect identical arch.
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
/tmp/ipython-input-1735790534.py in <cell line: 0>()
1 from transformers import pipeline
2
----> 3 ner = pipeline("ner", grouped_entities=True)
4 ner("My name is Sylvain and I work at Hugging Face in Brooklyn.")
2 frames
/usr/local/lib/python3.12/dist-packages/transformers/pipelines/base.py in __init__(self, model, tokenizer, feature_extractor, image_processor, processor, task, device, binary_output, **kwargs)
919 self._batch_size = kwargs.pop("batch_size", None)
920 self._num_workers = kwargs.pop("num_workers", None)
--> 921 self._preprocess_params, self._forward_params, self._postprocess_params = self._sanitize_parameters(**kwargs)
922
923 # In processor only mode, we can get the modality processors from the processor
TypeError: TokenClassificationPipeline._sanitize_parameters() got an unexpected keyword argument 'grouped_entities'
### Who can help?
_No response_
### Information
- [x] 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)
### Reproduction
Just run the script and in the cell no 17 the build fails
### Expected behavior
Updated cell:
from transformers import pipeline
ner = pipeline("ner")
ner("My name is Sylvain and I work at Hugging Face in Brooklyn.") | closed | completed | false | 7 | [
"Good First Issue",
"bug"
] | [] | 2026-02-15T19:15:34Z | 2026-03-23T17:00:09Z | 2026-02-24T16:07:22Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | khyatimirani | 29,675,392 | MDQ6VXNlcjI5Njc1Mzky | User | false |
huggingface/transformers | 3,962,963,557 | I_kwDOCUB6oc7sNgZl | 44,155 | https://github.com/huggingface/transformers/issues/44155 | https://api.github.com/repos/huggingface/transformers/issues/44155 | [AudioFlamingo3] Batched inference produces incorrect results due to embedding/token leak between tracks | ### System Info
- transformers version: 5.0.0
- Platform: Linux-5.14.0-284.30.1.el9_2.x86_64-x86_64-with-glibc2.39
- Python version: 3.12.3
- Huggingface_hub version: 1.4.1
- Safetensors version: 0.6.2
- Accelerate version: 1.11.0
- Accelerate config: not found
- DeepSpeed version: 0.18.1
- PyTorch version (accelerator?): 2.8.0+cu128 (CUDA)
- Using distributed or parallel set-up in script?: <fill in>
- Using GPU in script?: <fill in>
- GPU type: NVIDIA A100-SXM4-80GB
### Who can help?
audio models: @eustlb @ebezzam @vasqu
### Information
- [x] 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)
### Reproduction
Batched inference in `AudioFlamingo3ForConditionalGeneration` produces incorrect results. The same audio track receives completely different descriptions depending on other tracks in a batch.
**Root cause:** The processor and model compute audio embedding counts differently:
- Processor sums audio lengths across windows first, then applies downsampling formula
- Model applies downsampling formula per window, then sums
This causes a mismatch, and `masked_scatter` misaligns embeddings across items in a batch.
```python
"""
Reproduction script for AudioFlamingo3 batched inference bug.
"""
import torch
import numpy as np
import soundfile as sf
from transformers import AudioFlamingo3ForConditionalGeneration, AutoProcessor
# ============================================================
# Create TWO very different audio tracks
# ============================================================
def create_aggressive_audio(duration_sec, filepath):
"""Aggressive, noisy, distorted audio."""
sr = 16000
n_samples = int(sr * duration_sec)
t = np.linspace(0, duration_sec, n_samples, dtype=np.float32)
# Heavy distorted bass
bass = np.sin(2 * np.pi * 60 * t)
bass = np.clip(bass * 3, -1, 1)
# Noise
noise = np.random.randn(n_samples).astype(np.float32) * 0.3
# Fast harsh percussion
percussion = np.zeros(n_samples, dtype=np.float32)
for i in range(0, n_samples, int(sr * 0.125)):
decay = np.exp(-np.linspace(0, 10, min(int(sr*0.125), n_samples - i)))
percussion[i:i+len(decay)] += decay * 0.5
# Harsh high frequencies
harsh = np.sin(2 * np.pi * 3000 * t) * 0.2
audio = np.clip(0.4*bass + 0.3*noise + 0.2*percussion + 0.1*harsh, -0.95, 0.95)
sf.write(filepath, audio.astype(np.float32), sr)
return filepath
def create_upbeat_audio(duration_sec, filepath):
"""Upbeat, cheerful, melodic audio."""
sr = 16000
n_samples = int(sr * duration_sec)
t = np.linspace(0, duration_sec, n_samples, dtype=np.float32)
# Cheerful melody (major scale)
melody_freqs = [523, 587, 659, 784, 880] # C major pentatonic
melody = np.zeros(n_samples, dtype=np.float32)
note_len = int(sr * 0.25)
for i in range(0, n_samples, note_len):
freq = melody_freqs[(i // note_len) % len(melody_freqs)]
end = min(i + note_len, n_samples)
note_t = np.arange(end - i) / sr
melody[i:end] += np.sin(2 * np.pi * freq * note_t) * np.exp(-note_t * 3) * 0.4
# Bouncy bass
bass = np.sin(2 * np.pi * 150 * t) * 0.3 * (np.sin(2 * np.pi * 2 * t) > 0)
# Light beat
beat = np.zeros(n_samples, dtype=np.float32)
for i in range(0, n_samples, int(sr * 0.5)):
decay_len = min(int(sr * 0.05), n_samples - i)
beat[i:i+decay_len] += np.exp(-np.linspace(0, 20, decay_len)) * 0.3
audio = np.clip(melody + bass + beat, -0.95, 0.95)
sf.write(filepath, audio.astype(np.float32), sr)
return filepath
print("Creating test audio files...")
track_a = create_aggressive_audio(259.0, "/tmp/track_aggressive.wav")
track_b = create_upbeat_audio(239.0, "/tmp/track_upbeat.wav")
print("\nLoading model...")
processor = AutoProcessor.from_pretrained("nvidia/music-flamingo-hf")
model = AudioFlamingo3ForConditionalGeneration.from_pretrained(
"nvidia/music-flamingo-hf",
device_map="auto",
)
# ============================================================
# Helper function
# ============================================================
def run_batch(paths):
conversations = [
[{"role": "user", "content": [
{"type": "audio", "path": p},
{"type": "text", "text": "Describe this audio in detail. What mood and style does it have?"}
]}]
for p in paths
]
inputs = processor.apply_chat_template(
conversations, tokenize=True, add_generation_prompt=True, return_dict=True,
).to(model.device)
inputs["input_features"] = inputs["input_features"].to(dtype=model.dtype)
with torch.no_grad():
outputs = model.generate(**inputs, max_new_tokens=100)
return processor.batch_decode(
outputs[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True
)
def run_single(path):
"""Run inference on a single track (no batching = no contamination)."""
return run_batch([path])[0]
# ============================================================
# Test 1: Single track inference (CORRECT - no contamination)
# ============================================================
print("\n" + "=" * 70)
print("TEST 1: SINGLE TRACK INFERENCE (no batching)")
print("=" * 70)
result_a_single = run_single(track_a)
result_b_single = run_single(track_b)
print(f"\nTrack A alone:")
print(f" {result_a_single[:200]}...")
print(f"\nTrack B alone:")
print(f" {result_b_single[:200]}...")
# ============================================================
# Test 2: Batched inference (BUGGY - contamination occurs)
# ============================================================
print("\n" + "=" * 70)
print("TEST 2: BATCHED INFERENCE (both tracks together)")
print("=" * 70)
results_batched = run_batch([track_a, track_b])
print(f"\nTrack A in batch:")
print(f" {results_batched[0][:200]}...")
print(f"\nTrack B in batch:")
print(f" {results_batched[1][:200]}...")
# ============================================================
# Comparison
# ============================================================
print("\n" + "=" * 70)
print("COMPARISON: Track B ")
print("=" * 70)
print(f"\nWhen processed ALONE (correct):")
print(f" {result_b_single[:250]}...")
print(f"\nWhen processed in BATCH:")
print(f" {results_batched[1][:250]}...")
```
My output is:
```
======================================================================
COMPARISON: Track B
======================================================================
When processed ALONE (correct):
This track is a minimalist electronic piece that blends ambient and experimental styles. It features a steady, repetitive synth melody with a clean, digital timbre, set in a moderate tempo of 120 BPM and a 4/4 time signature. The production is sparse...
When processed in BATCH:
This track is a minimalist electronic piece that blends ambient and experimental styles, creating a calm, introspective atmosphere. It features a steady 4/4 beat at 120 BPM and is set in C major. The harmonic content is sparse, alternating between mo...
```
1. As we can see, there is a difference in the result.
2. The difference even more severe if we infer real music tracks from completely different genres like soft music and heavy. For the batch of two of them, the last one can be completely mislead by previous.
### Expected behavior
The outputs should be identical in all cases and not affected by other elements in a batch.
## Suggested Fix
The model's `get_audio_features` should group embeddings by sample and use the same formula as the processor:
```python
def get_audio_features(self, input_features, input_features_mask, windows_per_sample, **kwargs):
audio_output = self.audio_tower(input_features, input_features_mask=input_features_mask, return_dict=True, **kwargs)
audio_embeds = self.multi_modal_projector(audio_output.last_hidden_state)
post_lengths = (input_features_mask.sum(-1) - 2) // 2 + 1
all_sample_embeds = []
window_idx = 0
for n_windows in windows_per_sample:
sample_mask = input_features_mask[window_idx:window_idx + n_windows]
total_mask_sum = sample_mask.sum()
# Use processor's formula
conv_length = (total_mask_sum - 1) // 2 + 1
token_count = int(((conv_length - 2) // 2 + 1).item())
sample_embeds = audio_embeds[window_idx:window_idx + n_windows]
sample_post_lengths = post_lengths[window_idx:window_idx + n_windows]
max_len = sample_embeds.shape[1]
valid_mask = torch.arange(max_len, device=sample_post_lengths.device)[None, :] < sample_post_lengths[:, None]
sample_flat = sample_embeds[valid_mask][:token_count]
all_sample_embeds.append(sample_flat)
window_idx += n_windows
return torch.cat(all_sample_embeds, dim=0)
```
This requires passing `windows_per_sample` from the processor to the model, which may need API changes.
```python
# Constants from processor
SAMPLING_RATE = 16000
CHUNK_LENGTH = 30.0
MAX_AUDIO_LEN = 600
WINDOW_SIZE = int(SAMPLING_RATE * CHUNK_LENGTH)
MAX_WINDOWS = int(MAX_AUDIO_LEN // CHUNK_LENGTH)
def get_windows_per_sample(batch_paths):
"""Calculate number of windows for each audio file."""
windows_per_sample = []
for path in batch_paths:
audio, sr = librosa.load(path, sr=SAMPLING_RATE)
n_samples = len(audio)
n_windows = max(1, (n_samples + WINDOW_SIZE - 1) // WINDOW_SIZE)
if n_windows > MAX_WINDOWS:
n_windows = MAX_WINDOWS
windows_per_sample.append(n_windows)
return windows_per_sample
``` | open | null | false | 2 | [
"bug",
"Audio"
] | [] | 2026-02-19T13:20:07Z | 2026-03-22T08:04:17Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | IvanBarabanau | 34,709,713 | MDQ6VXNlcjM0NzA5NzEz | User | false |
huggingface/transformers | 3,967,173,673 | I_kwDOCUB6oc7sdkQp | 44,169 | https://github.com/huggingface/transformers/issues/44169 | https://api.github.com/repos/huggingface/transformers/issues/44169 | Need an example for FSDP + FP16 training | In my setup, I am trying to run FSDP with FP16 precision. Is there any limitation that I can not use FSDP with FP16 precision? How can I convert my existing code to FSDP for FP16 precision? I believe there is ShardedGradScaler from FSDP should be used. How is it different than normal GradScaler in terms of implementation? It will be great if someone share a concise example for this. | open | null | false | 2 | [] | [] | 2026-02-20T08:04:37Z | 2026-03-23T08:15:04Z | null | CONTRIBUTOR | null | 20260324T120231Z | 2026-03-24T12:02:31Z | quic-meetkuma | 200,747,495 | U_kgDOC_cp5w | User | false |
huggingface/transformers | 3,970,357,222 | I_kwDOCUB6oc7sptfm | 44,188 | https://github.com/huggingface/transformers/issues/44188 | https://api.github.com/repos/huggingface/transformers/issues/44188 | Diverging attention kernels due to `allow_is_bidirectional_skip` branching on torch.compile | ### System Info
Hi, while we were updating the PyTorch transformers pin to v5.2.0, our regression tests caught a numerics issue between eager and compiled, the difference is very substantial (3.3 vs the typical e-4 accepted difference). Digging into it: https://github.com/pytorch/pytorch/pull/175274#issuecomment-3930952666, we found the cause to be in these lines (added in https://github.com/huggingface/transformers/pull/41265):
https://github.com/huggingface/transformers/blob/147b7aa040812b079f467e777a2d2e1284167de0/src/transformers/masking_utils.py#L490-L491
We set `allow_is_bidirectional_skip=True` in a few places:
https://github.com/huggingface/transformers/blob/147b7aa040812b079f467e777a2d2e1284167de0/src/transformers/masking_utils.py#L996-L997
And in `_ignore_bidirectional_mask_sdpa`, we branch logic on whether we compile or not:
https://github.com/huggingface/transformers/blob/147b7aa040812b079f467e777a2d2e1284167de0/src/transformers/masking_utils.py#L324-L332
This issue was found on BERT but it seems like it would affect other models too.
We've also verified that removing the branching fixes the numerical difference. I'm creating this issue to ask about the best way forward here. From the PR that added it, it looks like this was necessary specifically for executorch, but the algorithm difference is also affected all other APIs that fall under `is_tracing` . Can we restrict the check?
### Who can help?
@vasqu @ArthurZucker @Cyrilvallez
### 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)
### Reproduction
I believe the description is enough, but I can provide a simpler repro on request
### Expected behavior
transformers users probably shouldn't run into large numeric differences when compiling, at least not by default | open | null | false | 8 | [
"bug"
] | [] | 2026-02-20T21:01:05Z | 2026-03-23T08:15:02Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | xmfan | 9,547,562 | MDQ6VXNlcjk1NDc1NjI= | User | false |
huggingface/transformers | 3,973,467,921 | I_kwDOCUB6oc7s1k8R | 44,208 | https://github.com/huggingface/transformers/issues/44208 | https://api.github.com/repos/huggingface/transformers/issues/44208 | request refund | I’m having a problem. I added my prepaid card, but the subscription was not accepted because the platform does not accept prepaid cards. However, the amount was deducted from my balance, and now I have neither the balance nor the Pro plan. I need a refund since I won’t be able to use the platform.
| closed | completed | false | 2 | [] | [] | 2026-02-22T03:42:03Z | 2026-03-24T13:31:02Z | 2026-03-24T13:31:02Z | NONE | null | 20260324T143453Z | 2026-03-24T14:34:53Z | storescienza-gif | 263,101,692 | U_kgDOD66c_A | User | false |
huggingface/transformers | 4,027,812,864 | I_kwDOCUB6oc7wE4wA | 44,462 | https://github.com/huggingface/transformers/issues/44462 | https://api.github.com/repos/huggingface/transformers/issues/44462 | AutoTokenizer ignores tokenizer.json from the repository | ### System Info
- `transformers` version: 5.3.0
- Python version: 3.10.12
- Huggingface_hub version: 1.5.0
### 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, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
`AutoTokenizer` doesn't load tokenizer based on `tokenizer.json` from the repository. Saving this tokenizer produces different `tokenizer.json` file.
```python
from transformers import AutoTokenizer
hf_tokenizer = AutoTokenizer.from_pretrained("deepseek-ai/deepseek-coder-6.7b-instruct")
hf_tokenizer.save_pretrained("hf_deepseek_tokenizer/")
```
[tokenizer_original.json](https://huggingface.co/deepseek-ai/deepseek-coder-6.7b-instruct/raw/main/tokenizer.json)
[tokenizer_saved.json](https://github.com/user-attachments/files/25766810/tokenizer_saved.json)
Original normalizer/pre-tokenizer:
```json
"normalizer": {
"type": "Sequence",
"normalizers": []
},
"pre_tokenizer": {
"type": "Sequence",
"pretokenizers": [
{
"type": "Split",
"pattern": {
"Regex": "[\r\n]"
},
"behavior": "Isolated",
"invert": false
},
{
"type": "Split",
"pattern": {
"Regex": "\\s?\\p{L}+"
},
"behavior": "Isolated",
"invert": false
},
{
"type": "Split",
"pattern": {
"Regex": "\\s?\\p{P}+"
},
"behavior": "Isolated",
"invert": false
},
{
"type": "Split",
"pattern": {
"Regex": "[一-龥ࠀ-一가-]+"
},
"behavior": "Isolated",
"invert": false
},
{
"type": "Digits",
"individual_digits": true
},
{
"type": "ByteLevel",
"add_prefix_space": false,
"trim_offsets": true,
"use_regex": false
}
]
}
```
Saved normalizer/pre-tokenizer:
```json
"normalizer": null,
"pre_tokenizer": {
"type": "Metaspace",
"replacement": "▁",
"prepend_scheme": "always",
"split": false
},
```
### Expected behavior
Original `tokenizer.json` should be used to instantiate the tokenizer. | closed | completed | false | 3 | [
"bug"
] | [] | 2026-03-05T12:04:31Z | 2026-03-20T12:37:12Z | 2026-03-20T12:37:12Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | apaniukov | 51,917,466 | MDQ6VXNlcjUxOTE3NDY2 | User | false |
huggingface/transformers | 4,032,973,273 | I_kwDOCUB6oc7wYknZ | 44,488 | https://github.com/huggingface/transformers/issues/44488 | https://api.github.com/repos/huggingface/transformers/issues/44488 | Current version also does not load "cjvt/sleng-bert" | ### System Info
broken config:
Python 3.13.5
tokenizers 0.22.2
transformers 5.2.0
torch 2.7.1+cu118
working config:
Python 3.13.5
tokenizers 0.22.1
transformers 4.57.1
torch 2.8.0+cu129
### Who can help?
@ArthurZucker @Cyrilvallez
### Information
- [x] The official example scripts
- [ ] My own modified scripts
### Tasks
- [x] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
```
>>> from transformers import AutoTokenizer
>>> bert_tokenizer = AutoTokenizer.from_pretrained("cjvt/sleng-bert")
Traceback (most recent call last):
File "<python-input-2>", line 1, in <module>
bert_tokenizer = AutoTokenizer.from_pretrained("cjvt/sleng-bert")
File "/nlp/scr/horatio/miniconda3/lib/python3.13/site-packages/transformers/models/auto/tokenization_auto.py", line 749, in from_pretrained
return tokenizer_class.from_pretrained(pretrained_model_name_or_path, *inputs, **kwargs)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/nlp/scr/horatio/miniconda3/lib/python3.13/site-packages/transformers/tokenization_utils_base.py", line 1721, in from_pretrained
return cls._from_pretrained(
~~~~~~~~~~~~~~~~~~~~^
resolved_vocab_files,
^^^^^^^^^^^^^^^^^^^^^
...<9 lines>...
**kwargs,
^^^^^^^^^
)
^
File "/nlp/scr/horatio/miniconda3/lib/python3.13/site-packages/transformers/tokenization_utils_base.py", line 1910, in _from_pretrained
tokenizer = cls(*init_inputs, **init_kwargs)
File "/nlp/scr/horatio/miniconda3/lib/python3.13/site-packages/transformers/models/camembert/tokenization_camembert.py", line 118, in __init__
unk_index = next((i for i, (tok, _) in enumerate(self._vocab) if tok == str(unk_token)), 0)
File "/nlp/scr/horatio/miniconda3/lib/python3.13/site-packages/transformers/models/camembert/tokenization_camembert.py", line 118, in <genexpr>
unk_index = next((i for i, (tok, _) in enumerate(self._vocab) if tok == str(unk_token)), 0)
^^^^^^^^
ValueError: too many values to unpack (expected 2)
```
### Expected behavior
Loading the model would be great! The older version of transformers works fine. | closed | completed | false | 14 | [
"bug"
] | [] | 2026-03-06T08:36:44Z | 2026-03-24T10:05:50Z | 2026-03-23T08:52:03Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | AngledLuffa | 3,411,033 | MDQ6VXNlcjM0MTEwMzM= | User | false |
huggingface/transformers | 4,040,512,605 | I_kwDOCUB6oc7w1VRd | 44,521 | https://github.com/huggingface/transformers/issues/44521 | https://api.github.com/repos/huggingface/transformers/issues/44521 | apply_chat_template returns all-zero assistant_masks for multimodal inputs | ### System Info
transformers==5.3.0
### Who can help?
@ArthurZucker and @itazap
### 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)
### Reproduction
```python
from transformers import AutoProcessor
messages = [
dict(
role="user",
content=[
dict(type="image", image="test.jpg"),
dict(type="text", text="Describe the image above."),
],
),
dict(
role="assistant",
content=[
dict(type="text", text="The image above shows a cat sitting on a table."),
],
),
]
processor = AutoProcessor.from_pretrained("Qwen/Qwen3.5-0.8B")
inputs = processor.apply_chat_template(
messages,
tokenize=True,
return_dict=True,
return_tensors="pt",
return_assistant_tokens_mask=True,
)
print(inputs["assistant_masks"])
```
### Actual behavior
assistant_masks is all zeros.
### Expected behavior
The tokens corresponding to the assistant response should be marked with 1 in assistant_masks.
### Suspected cause
I think the issue comes from a mismatch between generation_indices and the final tokenized prompt in multimodal cases.
In AutoProcessor.apply_chat_template, the assistant mask is computed roughly like this:
```python
prompt, generation_indices = render_jinja_template(
conversations=conversations,
chat_template=chat_template,
**template_kwargs,
**special_tokens_map,
)
...
out = self(
text=prompt,
images=batch_images if images_exist else None,
videos=batch_videos if videos_exist else None,
audio=batch_audios if batch_audios else None,
**kwargs,
)
...
offset_mapping = out.pop("offset_mapping")
input_ids = out["input_ids"]
for assistant_start_char, assistant_end_char in generation_indices[i]:
start_pos = bisect.bisect_left(offset_starts, assistant_start_char)
end_pos = bisect.bisect_left(offset_starts, assistant_end_char)
```
My understanding is:
- generation_indices is computed from the rendered text prompt returned by render_jinja_template
- but in multimodal processing, extra placeholder tokens such as <|image_pad|> are inserted later by the processor/tokenizer path
- therefore offset_mapping corresponds to the expanded multimodal text, while generation_indices still refers to the pre-expanded text
- this makes the character spans misaligned, so the assistant span lookup fails and assistant_masks ends up all zeros | open | null | false | 7 | [
"bug"
] | [] | 2026-03-08T05:03:12Z | 2026-03-24T14:46:59Z | null | NONE | null | 20260324T145640Z | 2026-03-24T14:56:40Z | renhouxing | 36,196,749 | MDQ6VXNlcjM2MTk2NzQ5 | User | false |
huggingface/transformers | 4,044,495,369 | I_kwDOCUB6oc7xEhoJ | 44,541 | https://github.com/huggingface/transformers/issues/44541 | https://api.github.com/repos/huggingface/transformers/issues/44541 | Can not deploy SFT Qwen3.5-9B model | ### System Info
I SFT Qwen3.5-9B model with transformers==5.2.0
### Reproduction
When I try to deploy my model with vllm==0.17.0, it reports:
TypeError: Invalid type of HuggingFace config. Expected type: <class 'vllm.transformers_utils.configs.qwen3_5.Qwen3_5Config'>, but found type: <class 'transformers.models.qwen3_5.configuration_qwen3_5.Qwen3_5TextConfig'>
I wonder how can I fix it.
### Expected behavior
vllm==0.17.0 requires transformers<5, but transformers>5 is required to support Qwen3.5, kind of strange. | open | null | false | 13 | [
"bug"
] | [] | 2026-03-09T09:39:19Z | 2026-03-23T12:16:54Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | zouYC2021 | 66,997,535 | MDQ6VXNlcjY2OTk3NTM1 | User | false |
huggingface/transformers | 4,051,220,122 | I_kwDOCUB6oc7xeLaa | 44,568 | https://github.com/huggingface/transformers/issues/44568 | https://api.github.com/repos/huggingface/transformers/issues/44568 | [BUG] add_special_tokens=True doesn't add BOS/EOS tokens for microsoft/mdeberta-v3-base tokenizer in transformers >=5.0 | ### System Info
## Version Details
- Working version: transformers==4.48.0
- Broken versions: transformers==5.0.0, 5.1.0, 5.2.0, 5.3.0
## Environment
- transformers: 5.2.0
- tokenizers: 0.22.2
- Python: 3.12
- Platform: Linux
### Who can help?
@ArthurZucker and @itazap
### 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)
### Reproduction
## Description
In transformers >=5.0, `add_special_tokens=True` doesn't add special tokens for `microsoft/mdeberta-v3-base` tokenizer. This is a regression from v4.x.
## Reproduction
```python
from transformers import AutoTokenizer
models = [
"microsoft/mdeberta-v3-base",
"FacebookAI/roberta-base",
"bert-base-uncased"
]
for model_name in models:
tokenizer = AutoTokenizer.from_pretrained(model_name)
result = tokenizer("hello", add_special_tokens=True)
print(f"{model_name}: input_ids={result['input_ids']}")
```
## Additional Notes
- The issue is MODEL-SPECIFIC, not general to all tokenizers
- Only microsoft/mdeberta-v3-base is affected
- The tokenizer has correct bos_token_id=1 and eos_token_id=2 values
- This appears to be related to DeBERTa v3's SentencePiece-based tokenizer and the v5 tokenizer redesign
## Expected Behavior
The behavior should be consistent across v4.x and v5.x for backward compatibility.
### Expected behavior
```
Expected (v4.48.0 - Working correctly)
- microsoft/mdeberta-v3-base: [1, 124394, 2] (CLS, hello, SEP)
- FacebookAI/roberta-base: [0, 42891, 2] (<s>, hello, </s>)
- bert-base-uncased: [101, 7592, 102] (CLS, hello, SEP)
Actual (v5.2.0 - Broken for mdeberta only)
- microsoft/mdeberta-v3-base: [124394] ← NO special tokens!
- FacebookAI/roberta-base: [0, 42891, 2] ← Works
- bert-base-uncased: [101, 7592, 102] ← Works
``` | closed | completed | false | 0 | [
"bug"
] | [] | 2026-03-10T11:43:59Z | 2026-03-24T09:40:46Z | 2026-03-24T09:40:46Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | Abdullahaml1 | 25,988,048 | MDQ6VXNlcjI1OTg4MDQ4 | User | false |
huggingface/transformers | 4,055,271,864 | I_kwDOCUB6oc7xtom4 | 44,589 | https://github.com/huggingface/transformers/issues/44589 | https://api.github.com/repos/huggingface/transformers/issues/44589 | TypeError: couldn't find storage object Float8_e4m3fnStorage | ### System Info
- `transformers` version: 5.3.0.dev0
- Platform: Linux-6.8.0-101-generic-x86_64-with-glibc2.35
- Python version: 3.12.13
- Huggingface_hub version: 1.6.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.10.0+cu128 (CUDA)
- Using distributed or parallel set-up in script?: <fill in>
- Using GPU in script?: <fill in>
- GPU type: NVIDIA A100-SXM4-80GB
### Who can help?
@CyrilVallez
### Information
- [x] The official example scripts
- [ ] 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)
### Reproduction
transformers serve --force-model Qwen/Qwen3.5-35B-A3B-FP8 --port 8000 --continuous-batching
```
File "/home/xinhe/auto-round/.venv/lib/python3.12/site-packages/transformers/cli/serve.py", line 474, in __init__
self.load_model_and_processor(model_id_and_revision)
File "/home/xinhe/auto-round/.venv/lib/python3.12/site-packages/transformers/cli/serve.py", line 1876, in load_model_and_processor
model, processor = self._load_model_and_data_processor(model_id_and_revision)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/xinhe/auto-round/.venv/lib/python3.12/site-packages/transformers/cli/serve.py", line 1850, in _load_model_and_data_processor
model = architecture.from_pretrained(model_id, **model_kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/xinhe/auto-round/.venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 4096, in from_pretrained
with ContextManagers(model_init_context):
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/xinhe/auto-round/.venv/lib/python3.12/site-packages/transformers/utils/generic.py", line 532, in __enter__
self.stack.enter_context(context_manager)
File "/home/xinhe/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 526, in enter_context
result = _enter(cm)
^^^^^^^^^^
File "/home/xinhe/.local/share/uv/python/cpython-3.12.13-linux-x86_64-gnu/lib/python3.12/contextlib.py", line 137, in __enter__
return next(self.gen)
^^^^^^^^^^^^^^
File "/home/xinhe/auto-round/.venv/lib/python3.12/site-packages/transformers/modeling_utils.py", line 240, in local_torch_dtype
torch.set_default_dtype(dtype)
File "/home/xinhe/auto-round/.venv/lib/python3.12/site-packages/torch/__init__.py", line 1358, in set_default_dtype
_C._set_default_dtype(d)
TypeError: couldn't find storage object Float8_e4m3fnStorage
```
### Expected behavior
The FP8 model should be loaded correctly. | closed | completed | false | 6 | [
"bug"
] | [] | 2026-03-11T02:16:53Z | 2026-03-20T13:12:32Z | 2026-03-20T13:12:32Z | CONTRIBUTOR | null | 20260324T120231Z | 2026-03-24T12:02:31Z | xin3he | 83,260,933 | MDQ6VXNlcjgzMjYwOTMz | User | false |
huggingface/transformers | 4,062,402,313 | I_kwDOCUB6oc7yI1cJ | 44,619 | https://github.com/huggingface/transformers/issues/44619 | https://api.github.com/repos/huggingface/transformers/issues/44619 | Plug model rule in development flow and extend it | This is a follow-up from https://github.com/huggingface/transformers/pull/44174
We're now plugging the tool into developer flow:
- add an opt-in github hook for checking the model
- automatically run `make check-model-rules` on PRs to generate reports
- add a ref to the CLI in the AI agents files
- fix models that are in allow lists and can be fixed
- modularize the code (add a `mlinter` sub package in `utils`)
| open | null | false | 0 | [] | [
"tarekziade"
] | 2026-03-12T06:00:58Z | 2026-03-20T15:45:01Z | null | MEMBER | null | 20260324T120231Z | 2026-03-24T12:02:31Z | tarekziade | 250,019 | MDQ6VXNlcjI1MDAxOQ== | User | false |
huggingface/transformers | 4,071,819,790 | I_kwDOCUB6oc7yswoO | 44,671 | https://github.com/huggingface/transformers/issues/44671 | https://api.github.com/repos/huggingface/transformers/issues/44671 | CamemBERT produces incorrect masked LM predictions in v5 | ### System Info
- `transformers` version: 5.3.0
- Platform: Linux-6.6.113+-x86_64-with-glibc2.35
- Python version: 3.12.12
- Huggingface_hub version: 1.6.0
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.10.0+cpu (NA)
- Using distributed or parallel set-up in script?: distributed
### Who can help?
@ArthurZucker @Cyrilvallez
### Information
- [x] 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)
### Reproduction
## Running the example from the documentation
https://huggingface.co/docs/transformers/model_doc/camembert?usage=Pipeline#camembert
```python
import torch
from transformers import pipeline
pipeline = pipeline("fill-mask", model="camembert-base", dtype=torch.float16, device=0)
pipeline("Le camembert est un délicieux fromage <mask>.")
```
v4
```text
[{'score': 0.181884765625,
'token': 4364,
'token_str': 'suisse',
'sequence': 'Le camembert est un délicieux fromage suisse .'},
{'score': 0.0943603515625,
'token': 430,
'token_str': 'français',
'sequence': 'Le camembert est un délicieux fromage français .'},
{'score': 0.049713134765625,
'token': 5941,
'token_str': 'italien',
'sequence': 'Le camembert est un délicieux fromage italien .'},
{'score': 0.036956787109375,
'token': 875,
'token_str': 'blanc',
'sequence': 'Le camembert est un délicieux fromage blanc .'},
{'score': 0.028778076171875,
'token': 18278,
'token_str': 'fondant',
'sequence': 'Le camembert est un délicieux fromage fondant .'}]
```
v5
```text
[{'score': 0.0001423358917236328,
'token': 14477,
'token_str': 'politique',
'sequence': 'Le camembert est un délicieux fromagepolitique .'},
{'score': 0.00013720989227294922,
'token': 16949,
'token_str': 'hâ',
'sequence': 'Le camembert est un délicieux fromagehâ .'},
{'score': 0.00011497735977172852,
'token': 7666,
'token_str': '<',
'sequence': 'Le camembert est un délicieux fromage< .'},
{'score': 0.00011020898818969727,
'token': 10494,
'token_str': 'entraîné',
'sequence': 'Le camembert est un délicieux fromage entraîné .'},
{'score': 0.00010907649993896484,
'token': 14341,
'token_str': 'améliore',
'sequence': 'Le camembert est un délicieux fromage améliore .'}]
```
## Simple masked token prediction test
```python
import random
import torch
from datasets import load_dataset
from tqdm import tqdm
from transformers import AutoTokenizer, AutoModelForMaskedLM, set_seed
set_seed(42)
tokenizer = AutoTokenizer.from_pretrained("camembert-base")
model = AutoModelForMaskedLM.from_pretrained("camembert-base", dtype="auto", device_map="auto", attn_implementation="sdpa")
dataset = load_dataset("allocine", split="test")
dataset = dataset.shuffle(seed=42)[:100]
num_correct = 0
num_total = 0
for text in tqdm(dataset["review"]):
input_ids = tokenizer(text, truncation=True, max_length=128)["input_ids"]
candidates = [
i for i, input_id in enumerate(input_ids)
if input_id not in tokenizer.all_special_ids
]
masked_pos = random.choice(candidates)
gold_id = input_ids[masked_pos]
masked_ids = input_ids[:]
masked_ids[masked_pos] = tokenizer.mask_token_id
with torch.no_grad():
output = model(torch.tensor([masked_ids]))
pred_id = output.logits[0, masked_pos].argmax(dim=-1)
num_correct += pred_id == gold_id
num_total += 1
print(f"\n\nnum_correct: {num_correct} / num_total: {num_total}")
```
v4
```text
num_correct: 43 / num_total: 100
```
v5
```text
num_correct: 0 / num_total: 100
```
### Expected behavior
CamemBERT masked language modeling should produce reasonable predictions as in transformers v4.
However, in transformers v5, the model produces clearly incorrect predictions and the simple masked token task accuracy drops from ~43% to 0%. | closed | completed | false | 5 | [
"bug"
] | [] | 2026-03-13T15:30:48Z | 2026-03-23T10:47:50Z | 2026-03-23T10:47:50Z | CONTRIBUTOR | null | 20260324T120231Z | 2026-03-24T12:02:31Z | math-hiyoko | 56,009,584 | MDQ6VXNlcjU2MDA5NTg0 | User | false |
huggingface/transformers | 4,077,031,787 | I_kwDOCUB6oc7zApFr | 44,717 | https://github.com/huggingface/transformers/issues/44717 | https://api.github.com/repos/huggingface/transformers/issues/44717 | Support packed sequences for linear attention models (i.e. Qwen3.5) | ### Feature request
Currently, packing does not seem supported for text-based datasets (https://github.com/unslothai/unsloth/issues/4160). It would be good to support this.
### Motivation
Without packing, my training runs are approximately 3-5x more expensive with the dataset that I'd like to use, and also suffer from overhead on very short sequences.
### Your contribution
I cannot help; I have no experience with deep learning. | open | null | false | 9 | [
"Feature request"
] | [] | 2026-03-14T23:22:19Z | 2026-03-23T16:21:08Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | kirawi | 67,773,714 | MDQ6VXNlcjY3NzczNzE0 | User | false |
huggingface/transformers | 4,080,877,033 | I_kwDOCUB6oc7zPT3p | 44,740 | https://github.com/huggingface/transformers/issues/44740 | https://api.github.com/repos/huggingface/transformers/issues/44740 | Run model linter on new models additions | - detect PRs with new models
- add a `new model` label
- run the linter and when failing comment the PR with errors | open | null | false | 1 | [] | [
"tarekziade"
] | 2026-03-16T08:04:16Z | 2026-03-23T09:07:09Z | null | MEMBER | null | 20260324T120231Z | 2026-03-24T12:02:31Z | tarekziade | 250,019 | MDQ6VXNlcjI1MDAxOQ== | User | false |
huggingface/transformers | 4,081,656,083 | I_kwDOCUB6oc7zSSET | 44,748 | https://github.com/huggingface/transformers/issues/44748 | https://api.github.com/repos/huggingface/transformers/issues/44748 | [Neuron] Auto-select StaticCache when device is Neuron | ## Context
On Neuron devices, `StaticCache` is required for correct generation — dynamic tensor shapes trigger per-step recompilations. Currently, users must explicitly pass `past_key_values=StaticCache(...)` or `cache_implementation="static"` to `model.generate()`. Ideally, `model.generate()` should auto-select `StaticCache` when running on Neuron, providing zero-config DX.
**Depends on:** #44742 (StaticCache-friendly `_sample` must exist first — otherwise auto-selecting StaticCache still hits the dynamic-shape `_sample` path)
## Proposed change
In `_prepare_cache_for_generation` (around line 1879 of `src/transformers/generation/utils.py`), when no cache is specified:
```python
if generation_config.cache_implementation is None and self.device.type == "neuron":
generation_config.cache_implementation = "static"
```
**Precedent:** The generation code already has device-aware logic:
- Line 2021: `self.device.type in ["cuda", "xpu"]` for compile criteria
- Line 1870: Forces `"dynamic_full"` for assisted generation
Device-aware cache selection fits the existing pattern.
## Expected behavior
```python
# On Neuron — auto-selects StaticCache, uses static-shape _sample path:
model.generate(input_ids)
# On CUDA — unchanged, uses DynamicCache + standard _sample:
model.generate(input_ids)
# Explicit override still works on any device:
model.generate(input_ids, cache_implementation="static")
model.generate(input_ids, past_key_values=my_static_cache)
```
## Related
- Parent: #44741
- Depends on: #44742 | open | null | false | 1 | [] | [] | 2026-03-16T10:38:15Z | 2026-03-21T18:19:03Z | null | MEMBER | null | 20260324T120231Z | 2026-03-24T12:02:31Z | dacorvo | 1,910,518 | MDQ6VXNlcjE5MTA1MTg= | User | false |
huggingface/transformers | 4,081,994,049 | I_kwDOCUB6oc7zTklB | 44,749 | https://github.com/huggingface/transformers/issues/44749 | https://api.github.com/repos/huggingface/transformers/issues/44749 | Transformer 从4.57.3 升级到5.3.0 后过滤数据时长变慢十倍以上 | ### System Info
H20
### 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 details below)
### Reproduction
过滤函数实现：对应链接 https://github.com/verl-project/verl/blob/main/verl/utils/dataset/rl_dataset.py#L241
```
def doc2len(doc) -> int:
try:
apply_kwargs = dict(**self.apply_chat_template_kwargs)
if self.tool_schemas is not None:
apply_kwargs["tools"] = self.tool_schemas
# Keep explicit tokenization to avoid transformers version default changes.
apply_kwargs.pop("tokenize", None)
apply_kwargs.pop("return_dict", None)
apply_kwargs.pop("return_tensors", None)
tokenized_prompt = tokenizer.apply_chat_template(
doc[prompt_key], add_generation_prompt=True, tokenize=True, **apply_kwargs
)
return len(normalize_token_ids(tokenized_prompt))
except Exception:
print("Error processing one of the samples, skipping...")
traceback.print_exc()
return self.max_prompt_length + 1
dataframe = dataframe.filter(
lambda doc: doc2len(doc) <= self.max_prompt_length,
num_proc=self.num_workers,
desc=f"Filtering prompts longer than {self.max_prompt_length} tokens",
)
```
使用transformer 4.57.3 运行这段code大概十几分钟，如果升级到5.3.0 需要约两个小时；辛苦帮忙看看 模型=qwen3.5-35b-A22 数据集是aime-2024
4.57.3 结果
<img width="1818" height="300" alt="Image" src="https://github.com/user-attachments/assets/df875792-23ae-4efe-9f42-c11de9516e5a" />
5.3.0 结果
<img width="1506" height="236" alt="Image" src="https://github.com/user-attachments/assets/d2b7f182-0337-4cb0-a32e-d34cfd78f43a" />
### Expected behavior
希望transformer 升级后，其tokenizer 的速度没有变化 | open | null | false | 7 | [
"bug"
] | [] | 2026-03-16T11:48:25Z | 2026-03-23T08:28:15Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | chenjiaoAngel | 38,650,344 | MDQ6VXNlcjM4NjUwMzQ0 | User | false |
huggingface/transformers | 4,082,686,570 | I_kwDOCUB6oc7zWNpq | 44,754 | https://github.com/huggingface/transformers/issues/44754 | https://api.github.com/repos/huggingface/transformers/issues/44754 | I really need help regarding to meta device issue | I installed TRELLIS 2 and getting the error for meta device, I followed the instructions here for installing this release of transformers, but when I launch TRELLIS 2 I still get the same error. Can someone please help me Fix this step by step because I spend 5 days trying without any solution | open | null | false | 1 | [] | [] | 2026-03-16T14:03:28Z | 2026-03-23T09:44:39Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | Hany138078 | 230,026,885 | U_kgDODbXuhQ | User | false |
huggingface/transformers | 4,093,576,102 | I_kwDOCUB6oc7z_wOm | 44,811 | https://github.com/huggingface/transformers/issues/44811 | https://api.github.com/repos/huggingface/transformers/issues/44811 | Whisper processor.batch_decode() function ignoring skip_special_tokens params | ### System Info
- `transformers` version: 4.57.6
- Platform: Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.35
- Python version: 3.10.13
- Huggingface_hub version: 0.36.2
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.10.0+cu130 (CUDA)
- Tensorflow version (GPU?): not installed (NA)
- Flax version (CPU?/GPU?/TPU?): not installed (NA)
- Jax version: not installed
- JaxLib version: not installed
- Using distributed or parallel set-up in script?: No
- Using GPU in script?: Yes
- GPU type: NVIDIA GeForce RTX 3080 Ti Laptop GPU
### Who can help?
@ArthurZucker
### Information
- [x] The official example scripts
- [ ] My own modified scripts
### Tasks
- [x] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
Running the code provided [here](https://huggingface.co/openai/whisper-small), it seems that _skip_special_tokens_ param is being ignored.
```
from transformers import WhisperProcessor, WhisperForConditionalGeneration
from datasets import load_dataset
# load model and processor
processor = WhisperProcessor.from_pretrained("openai/whisper-small")
model = WhisperForConditionalGeneration.from_pretrained("openai/whisper-small")
model.config.forced_decoder_ids = None
# load dummy dataset and read audio files
ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
sample = ds[0]["audio"]
input_features = processor(sample["array"], sampling_rate=sample["sampling_rate"], return_tensors="pt").input_features
# generate token ids
predicted_ids = model.generate(input_features)
# decode token ids to text
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=False)
print("[Skip special tokens=False] ", transcription)
transcription = processor.batch_decode(predicted_ids, skip_special_tokens=True)
print("[Skip special tokens=True] ", transcription)
```
The output is the following:
_[Skip special tokens=False] [' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.']
[Skip special tokens=True] [' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.']_
If however a dict output is required by running the following code:
```
# generate token ids
predicted_ids = model.generate(input_features, return_dict_in_generate=True)
# decode token ids to text
transcription = processor.batch_decode(predicted_ids.sequences, skip_special_tokens=False)
print("[Skip special tokens=False] ", transcription)
transcription = processor.batch_decode(predicted_ids.sequences, skip_special_tokens=True)
print("[Skip special tokens=True] ", transcription)
```
The output is (correctly) the following:
_[Skip special tokens=False] ['<|startoftranscript|><|en|><|transcribe|><|notimestamps|> Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.<|endoftext|>']
[Skip special tokens=True] [' Mr. Quilter is the apostle of the middle classes, and we are glad to welcome his gospel.']_
### Expected behavior
The _processor.batch_decode()_ function should output also special tokens if _skip_special_tokens=False_ is passed. | closed | completed | false | 5 | [
"bug"
] | [] | 2026-03-18T08:30:17Z | 2026-03-20T10:59:59Z | 2026-03-20T10:59:59Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | cfasana | 143,723,410 | U_kgDOCJELkg | User | false |
huggingface/transformers | 4,094,388,173 | I_kwDOCUB6oc70C2fN | 44,821 | https://github.com/huggingface/transformers/issues/44821 | https://api.github.com/repos/huggingface/transformers/issues/44821 | Unable to load `AutoImageProcessor` from URL | ### System Info
<details><summary>Versions</summary>
<p>
- `transformers` version: 5.3.0
- Platform: Linux-6.8.0-106-generic-x86_64-with-glibc2.39
- Python version: 3.12.3
- Huggingface_hub version: 1.7.1
- Safetensors version: 0.7.0
- Accelerate version: not installed
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.10.0+cu128 (NA)
</p>
</details>
### Bug
I am trying to load a config from URL, to instatiate the AutoImageProcessor. This worked before, but not anymore with new versions of transformers.
Bug starts occuring with `transformers>=5.3.0`. It worked with `transformers==4.57.6`
To reproduce;
```py
import transformers
url = "https://huggingface.co/jinfengxie/BFMS_1014/raw/main/config.json"
transformers.AutoImageProcessor.from_pretrained(url) # fails on >=5.0, succeeds pre-5.0
local_file = "/path/to/config.json" # download the file from the URL above
transformers.AutoImageProcessor.from_pretrained(local_file) # succeeds on v4 and v5
```
<details><summary>traceback</summary>
<p>
```
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/home/bart/git/streetscapes/.venv/lib/python3.12/site-packages/transformers/models/auto/image_processing_auto.py", line 514, in from_pretrained
raise initial_exception
File "/home/bart/git/streetscapes/.venv/lib/python3.12/site-packages/transformers/models/auto/image_processing_auto.py", line 501, in from_pretrained
config_dict, _ = ImageProcessingMixin.get_image_processor_dict(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/bart/git/streetscapes/.venv/lib/python3.12/site-packages/transformers/image_processing_base.py", line 282, in get_image_processor_dict
resolved_processor_file = cached_file(
^^^^^^^^^^^^
File "/home/bart/git/streetscapes/.venv/lib/python3.12/site-packages/transformers/utils/hub.py", line 277, in cached_file
file = cached_files(path_or_repo_id=path_or_repo_id, filenames=[filename], **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/home/bart/git/streetscapes/.venv/lib/python3.12/site-packages/transformers/utils/hub.py", line 469, in cached_files
raise OSError(f"{e}") from e
OSError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': 'https://huggingface.co/jinfengxie/BFMS_1014/raw/main/config.json'. Use `repo_type` argument if needed.
```
</p>
</details>
| closed | completed | false | 6 | [
"bug"
] | [] | 2026-03-18T11:08:09Z | 2026-03-23T13:13:38Z | 2026-03-23T13:10:59Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | BSchilperoort | 12,114,825 | MDQ6VXNlcjEyMTE0ODI1 | User | false |
huggingface/transformers | 4,098,836,913 | I_kwDOCUB6oc70T0mx | 44,841 | https://github.com/huggingface/transformers/issues/44841 | https://api.github.com/repos/huggingface/transformers/issues/44841 | Processor fails for mistralai/Voxtral-Mini-3B-2507 | ### System Info
I am trying to run inference using `mistralai/Voxtral-Mini-3B-2507` on an audio (`np.ndarray`). On loading the processor using `processor = transformers.AutoProcessor.from_pretrained(MODEL, trust_remote_code=True)`, I am getting the following error:
```
/usr/local/lib/python3.12/dist-packages/huggingface_hub/utils/_auth.py:94: UserWarning:
The secret `HF_TOKEN` does not exist in your Colab secrets.
To authenticate with the Hugging Face Hub, create a token in your settings tab (https://huggingface.co/settings/tokens), set it as secret in your Google Colab and restart your session.
You will be able to reuse this secret in all of your notebooks.
Please note that authentication is recommended but still optional to access public models or datasets.
warnings.warn(
Converting tekken.json to tokenizer.json: 100%|██████████| 150000/150000 [00:02<00:00, 69142.59it/s]
---------------------------------------------------------------------------
Exception Traceback (most recent call last)
[/tmp/ipykernel_7845/2744582412.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in <cell line: 0>()
12 # else:
13 # processor = MODEL_DICT[MODEL]['processor'].from_pretrained(MODEL, trust_remote_code=True)
---> 14 processor = MODEL_DICT[MODEL]['processor'].from_pretrained(MODEL, trust_remote_code=True)
13 frames[/usr/local/lib/python3.12/dist-packages/transformers/models/auto/processing_auto.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in from_pretrained(cls, pretrained_model_name_or_path, **kwargs)
409 )
410 elif processor_class is not None:
--> 411 return processor_class.from_pretrained(
412 pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
413 )
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in from_pretrained(cls, pretrained_model_name_or_path, cache_dir, force_download, local_files_only, token, revision, **kwargs)
1402 processor_dict, instantiation_kwargs = cls.get_processor_dict(pretrained_model_name_or_path, **kwargs)
1403 args = cls._get_arguments_from_pretrained(pretrained_model_name_or_path, processor_dict, **kwargs)
-> 1404 return cls.from_args_and_dict(args, processor_dict, **instantiation_kwargs)
1405
1406 @classmethod
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in from_args_and_dict(cls, args, processor_dict, **kwargs)
1171 processor = cls(*args, **valid_kwargs)
1172
-> 1173 logger.info(f"Processor {processor}")
1174 if return_unused_kwargs:
1175 return processor, unused_kwargs
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in __repr__(self)
777 attributes_repr = [f"- {name}: {repr(getattr(self, name))}" for name in self.get_attributes()]
778 attributes_repr = "\n".join(attributes_repr)
--> 779 return f"{self.__class__.__name__}:\n{attributes_repr}\n\n{self.to_json_string()}"
780
781 def save_pretrained(self, save_directory, push_to_hub: bool = False, **kwargs):
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in to_json_string(self)
759 `str`: String containing all the attributes that make up this feature_extractor instance in JSON format.
760 """
--> 761 dictionary = self.to_dict()
762
763 return json.dumps(dictionary, indent=2, sort_keys=True) + "\n"
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in to_dict(self)
697 `dict[str, Any]`: Dictionary of all the attributes that make up this processor instance.
698 """
--> 699 output = copy.deepcopy(self.__dict__)
700
701 # Get the kwargs in `__init__`.
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in deepcopy(x, memo, _nil)
134 copier = _deepcopy_dispatch.get(cls)
135 if copier is not None:
--> 136 y = copier(x, memo)
137 else:
138 if issubclass(cls, type):
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in _deepcopy_dict(x, memo, deepcopy)
219 memo[id(x)] = y
220 for key, value in x.items():
--> 221 y[deepcopy(key, memo)] = deepcopy(value, memo)
222 return y
223 d[dict] = _deepcopy_dict
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in deepcopy(x, memo, _nil)
160 y = x
161 else:
--> 162 y = _reconstruct(x, memo, *rv)
163
164 # If is its own copy, don't memoize.
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in _reconstruct(x, memo, func, args, state, listiter, dictiter, deepcopy)
257 if state is not None:
258 if deep:
--> 259 state = deepcopy(state, memo)
260 if hasattr(y, '__setstate__'):
261 y.__setstate__(state)
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in deepcopy(x, memo, _nil)
134 copier = _deepcopy_dispatch.get(cls)
135 if copier is not None:
--> 136 y = copier(x, memo)
137 else:
138 if issubclass(cls, type):
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in _deepcopy_dict(x, memo, deepcopy)
219 memo[id(x)] = y
220 for key, value in x.items():
--> 221 y[deepcopy(key, memo)] = deepcopy(value, memo)
222 return y
223 d[dict] = _deepcopy_dict
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in deepcopy(x, memo, _nil)
160 y = x
161 else:
--> 162 y = _reconstruct(x, memo, *rv)
163
164 # If is its own copy, don't memoize.
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in _reconstruct(x, memo, func, args, state, listiter, dictiter, deepcopy)
259 state = deepcopy(state, memo)
260 if hasattr(y, '__setstate__'):
--> 261 y.__setstate__(state)
262 else:
263 if isinstance(state, tuple) and len(state) == 2:
Exception: Error while attempting to unpickle Tokenizer: Token `Ġ<SPECIAL_116>` out of vocabulary at line 1 column 9601910
```
I tested out the following standalone code provided in the model card of `mistralai/Voxtral-Mini-3B-2507` but it too does not work:
```
from transformers import VoxtralForConditionalGeneration, AutoProcessor
import torch
device = "cuda"
repo_id = "mistralai/Voxtral-Mini-3B-2507"
processor = AutoProcessor.from_pretrained(repo_id)
model = VoxtralForConditionalGeneration.from_pretrained(repo_id, torch_dtype=torch.bfloat16, device_map=device)
conversation = [
{
"role": "user",
"content": [
{
"type": "audio",
"path": "https://huggingface.co/datasets/hf-internal-testing/dummy-audio-samples/resolve/main/winning_call.mp3",
},
],
}
]
inputs = processor.apply_chat_template(conversation)
inputs = inputs.to(device, dtype=torch.bfloat16)
outputs = model.generate(**inputs, max_new_tokens=500)
decoded_outputs = processor.batch_decode(outputs[:, inputs.input_ids.shape[1]:], skip_special_tokens=True)
print("\nGenerated response:")
print("=" * 80)
print(decoded_outputs[0])
print("=" * 80)
```
Error given by above script:
```
Converting tekken.json to tokenizer.json: 100%|██████████| 150000/150000 [00:01<00:00, 116597.64it/s]
---------------------------------------------------------------------------
Exception Traceback (most recent call last)
[/tmp/ipykernel_7845/3351359521.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in <cell line: 0>()
5 repo_id = "mistralai/Voxtral-Mini-3B-2507"
6
----> 7 processor = AutoProcessor.from_pretrained(repo_id)
8 model = VoxtralForConditionalGeneration.from_pretrained(repo_id, torch_dtype=torch.bfloat16, device_map=device)
9
13 frames[/usr/local/lib/python3.12/dist-packages/transformers/models/auto/processing_auto.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in from_pretrained(cls, pretrained_model_name_or_path, **kwargs)
409 )
410 elif processor_class is not None:
--> 411 return processor_class.from_pretrained(
412 pretrained_model_name_or_path, trust_remote_code=trust_remote_code, **kwargs
413 )
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in from_pretrained(cls, pretrained_model_name_or_path, cache_dir, force_download, local_files_only, token, revision, **kwargs)
1402 processor_dict, instantiation_kwargs = cls.get_processor_dict(pretrained_model_name_or_path, **kwargs)
1403 args = cls._get_arguments_from_pretrained(pretrained_model_name_or_path, processor_dict, **kwargs)
-> 1404 return cls.from_args_and_dict(args, processor_dict, **instantiation_kwargs)
1405
1406 @classmethod
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in from_args_and_dict(cls, args, processor_dict, **kwargs)
1171 processor = cls(*args, **valid_kwargs)
1172
-> 1173 logger.info(f"Processor {processor}")
1174 if return_unused_kwargs:
1175 return processor, unused_kwargs
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in __repr__(self)
777 attributes_repr = [f"- {name}: {repr(getattr(self, name))}" for name in self.get_attributes()]
778 attributes_repr = "\n".join(attributes_repr)
--> 779 return f"{self.__class__.__name__}:\n{attributes_repr}\n\n{self.to_json_string()}"
780
781 def save_pretrained(self, save_directory, push_to_hub: bool = False, **kwargs):
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in to_json_string(self)
759 `str`: String containing all the attributes that make up this feature_extractor instance in JSON format.
760 """
--> 761 dictionary = self.to_dict()
762
763 return json.dumps(dictionary, indent=2, sort_keys=True) + "\n"
[/usr/local/lib/python3.12/dist-packages/transformers/processing_utils.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in to_dict(self)
697 `dict[str, Any]`: Dictionary of all the attributes that make up this processor instance.
698 """
--> 699 output = copy.deepcopy(self.__dict__)
700
701 # Get the kwargs in `__init__`.
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in deepcopy(x, memo, _nil)
134 copier = _deepcopy_dispatch.get(cls)
135 if copier is not None:
--> 136 y = copier(x, memo)
137 else:
138 if issubclass(cls, type):
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in _deepcopy_dict(x, memo, deepcopy)
219 memo[id(x)] = y
220 for key, value in x.items():
--> 221 y[deepcopy(key, memo)] = deepcopy(value, memo)
222 return y
223 d[dict] = _deepcopy_dict
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in deepcopy(x, memo, _nil)
160 y = x
161 else:
--> 162 y = _reconstruct(x, memo, *rv)
163
164 # If is its own copy, don't memoize.
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in _reconstruct(x, memo, func, args, state, listiter, dictiter, deepcopy)
257 if state is not None:
258 if deep:
--> 259 state = deepcopy(state, memo)
260 if hasattr(y, '__setstate__'):
261 y.__setstate__(state)
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in deepcopy(x, memo, _nil)
134 copier = _deepcopy_dispatch.get(cls)
135 if copier is not None:
--> 136 y = copier(x, memo)
137 else:
138 if issubclass(cls, type):
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in _deepcopy_dict(x, memo, deepcopy)
219 memo[id(x)] = y
220 for key, value in x.items():
--> 221 y[deepcopy(key, memo)] = deepcopy(value, memo)
222 return y
223 d[dict] = _deepcopy_dict
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in deepcopy(x, memo, _nil)
160 y = x
161 else:
--> 162 y = _reconstruct(x, memo, *rv)
163
164 # If is its own copy, don't memoize.
[/usr/lib/python3.12/copy.py](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD#) in _reconstruct(x, memo, func, args, state, listiter, dictiter, deepcopy)
259 state = deepcopy(state, memo)
260 if hasattr(y, '__setstate__'):
--> 261 y.__setstate__(state)
262 else:
263 if isinstance(state, tuple) and len(state) == 2:
Exception: Error while attempting to unpickle Tokenizer: Token `<SPECIAL_32><SPECIAL_32>` out of vocabulary at line 1 column 9847997
```
There are many issues with a lot of other APIs in `transformers==5.3.0`. Please resolve these issues ASAP.
### Who can help?
@eustlb @ebezzam @vasqu @Cyrilvallez
### 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
Steps to reproduce the issue:
1. Connect to the correct GPU runtime (I tried on T4 GPU).
2. Upload a valid audio file with extension `.WAV`.
3. Ensure `MODEL = "mistralai/Voxtral-Mini-3B-2507"` and `AUDIO_PATH = "<valid audio path>.WAV"`.
4. Run the notebook at [https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD?usp=sharing](https://colab.research.google.com/drive/18tlO4aFg0RxV7hcBf04abb1wKW-B3zbD?usp=sharing).
[EMSCall (1).ipynb](https://github.com/user-attachments/files/26104503/EMSCall.1.ipynb)
### Expected behavior
Error on loading of `processor` for `mistralai/Voxtral-Mini-3B-2507` model. | closed | completed | false | 3 | [
"bug"
] | [] | 2026-03-19T02:26:39Z | 2026-03-24T10:07:36Z | 2026-03-24T10:07:36Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | BhavyaShah1234 | 62,424,200 | MDQ6VXNlcjYyNDI0MjAw | User | false |
huggingface/transformers | 4,099,433,723 | I_kwDOCUB6oc70WGT7 | 44,843 | https://github.com/huggingface/transformers/issues/44843 | https://api.github.com/repos/huggingface/transformers/issues/44843 | AutoTokenizer.from_pretrained calls model_info() unconditionally in _patch_mistral_regex, breaks HF_HUB_OFFLINE mode | ### System Info
- `transformers` version: 4.57.3
- `huggingface_hub` version: 0.36.2
- Python: 3.12
- OS: Linux (Ubuntu 24.04, inside NVIDIA container)
### Who can help?
@ArthurZucker @itazap
### Regression introduced in
PR #42389 (`[Mistral Tokenizers] Fix tokenizer detection`), included in v4.57.2 → v4.57.3.
### Information
- [x] The official example scripts
- [x] My own modified scripts
### Tasks
- [x] An officially supported task in the `examples` folder
- [ ] My own task or dataset (give details below)
### Reproduction
```python
import os
from huggingface_hub import snapshot_download
# Step 1: Pre-download a non-Mistral model
snapshot_download("Qwen/Qwen3-0.6B")
# Step 2: Enable offline mode
os.environ["HF_HUB_OFFLINE"] = "1"
# Step 3: Load tokenizer — crashes even though model is fully cached
from transformers import AutoTokenizer
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
```
This raises:
```
huggingface_hub.errors.OfflineModeIsEnabled: Cannot reach
https://huggingface.co/api/models/Qwen/Qwen3-0.6B: offline mode is enabled.
```
Full traceback:
```
transformers/models/auto/tokenization_auto.py:1156, in from_pretrained
transformers/tokenization_utils_base.py:2113, in from_pretrained
transformers/tokenization_utils_base.py:2395, in _from_pretrained
transformers/tokenization_utils_base.py:2438, in _patch_mistral_regex
transformers/tokenization_utils_base.py:2432, in is_base_mistral
huggingface_hub/hf_api.py:2660, in model_info
→ OfflineModeIsEnabled
```
### Expected behavior
`AutoTokenizer.from_pretrained()` should work in offline mode (`HF_HUB_OFFLINE=1`) when the model is already cached locally. This worked in transformers 4.57.1.
### Root cause
In `tokenization_utils_base.py`, `_patch_mistral_regex` defines `is_base_mistral()` which calls `huggingface_hub.model_info()` — an API call — for **every** model, not just Mistral models. The call happens before any cache check or `local_files_only` guard:
```python
# tokenization_utils_base.py, _patch_mistral_regex
def is_base_mistral(model_id: str) -> bool:
model = model_info(model_id) # <-- unconditional API call
...
if _is_local or is_base_mistral(pretrained_model_name_or_path): # <-- called for ALL non-local models
```
The `local_files_only` parameter is passed to `_patch_mistral_regex` but is never used to guard the `is_base_mistral()` call.
### Suggested fix
The `is_base_mistral()` call should either:
1. Be wrapped in a try/except that catches `OfflineModeIsEnabled` and returns `False` (safe default — if we can't reach the API, assume it's not a Mistral model), or
2. Be skipped when `local_files_only=True` or `HF_HUB_OFFLINE=1` is set
### Impact
This breaks any CI/CD pipeline or air-gapped environment that:
1. Pre-downloads models with `snapshot_download()`
2. Sets `HF_HUB_OFFLINE=1` to prevent network access during test execution
3. Loads tokenizers via `AutoTokenizer.from_pretrained()`
This pattern is common in ML CI pipelines (we hit this in NVIDIA Dynamo's CI with TensorRT-LLM). | open | null | false | 3 | [] | [] | 2026-03-19T05:36:56Z | 2026-03-23T11:15:07Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | nv-yna | 248,773,860 | U_kgDODtP85A | User | false |
huggingface/transformers | 4,100,120,851 | I_kwDOCUB6oc70YuET | 44,849 | https://github.com/huggingface/transformers/issues/44849 | https://api.github.com/repos/huggingface/transformers/issues/44849 | Transformers Qwen3.5 had a bug when set output_hidden_states=True | ### System Info
Version: 5.2.0
in qwen3.5
outputs = model_wrapper.generate(**inputs, output_hidden_states=True)
outpus something like this:
```
><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|image_pad|><|vision_end|>请详细描述这张图片的内容。<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n这张!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!']
```
ignore the `output_hideen_states` params, normal
```
on_end|>请详细描述这张图片的内容。<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n这张图片是一张包含表格的截图，内容是一份关于不同模型在多个任务上表现的数据对比表。\n\n**整体布局：**\n图片的主体是一个表格，表格的标题为“Model”，列出了多个模型名称，以及它们在不同任务上的平均长度（Avg. Length）、任务1、任务2、任务3、任务4和任务5的得分百分比。表格下方还有一行说明文字。\n\n**表格内容详情：**\n表格共有8行数据，对应8个不同的模型。\n\n- **第一行：**\n - **Model:** `PI0*`\n - **Avg. Length:** 2.954\n - **Task 1:** 84.8%\n - **Task 2:** 70.4%\n - **Task 3:** 55.9%\n - **Task 4:** 46.6%\n - **Task 5:** 37.7%\n\n- **第二行：**\n - **Model:** `PI0.5*`\n - **Avg. Length:** 3.885\n - **Task 1:** 92.5%\n - **Task 2:** 84.0%\n - **Task 3:** 76.6%\n - **Task 4:** 71.0%\n - **Task 5:** 64.4%\n\n- **第三行：**\n - **Model:** `qwenpi (qwen2.5-vl-3B-instruct-action)`\n - **Avg. Length:** 3.5']
```
### Who can help?
w
### 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
f
### Expected behavior
rg | open | null | false | 4 | [
"bug"
] | [] | 2026-03-19T08:27:35Z | 2026-03-20T01:27:23Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | lucasjinreal | 21,303,438 | MDQ6VXNlcjIxMzAzNDM4 | User | false |
huggingface/transformers | 4,100,959,949 | I_kwDOCUB6oc70b67N | 44,855 | https://github.com/huggingface/transformers/issues/44855 | https://api.github.com/repos/huggingface/transformers/issues/44855 | IndentationError when importing DebertaV2Model on Python 3.13 - @torch.jit.script fails to parse function with comment between decorator and def | ## Description
Importing `DebertaV2Model` from `transformers` (or any library that depends on it, such as `gliner`) raises an `IndentationError` on Python 3.13. The error originates in `torch.jit.script` when it attempts to re-parse the source of a JIT-scripted function that has a comment placed between the `@torch.jit.script` decorator and the `def` statement.
## Root Cause
In `modeling_deberta_v2.py`, several functions are decorated with `@torch.jit.script` with a comment in between:
```python
@torch.jit.script
# Copied from transformers.models.deberta.modeling_deberta.c2p_dynamic_expand
def c2p_dynamic_expand(c2p_pos, query_layer, relative_pos):
return c2p_pos.expand([...])
```
When `@torch.jit.script` is applied, PyTorch's JIT frontend (`torch/_sources.py`) calls `inspect.getsource()` on the function, dedents the source, and passes it to `ast.parse()`. On Python 3.13, the comment between the decorator and `def` is included in the extracted snippet. After dedenting, the snippet becomes:
```
@torch.jit.script
# Copied from ...
def c2p_dynamic_expand(...):
return ...
```
Python 3.13's `ast.parse` (backed by a stricter parser) then fails with `IndentationError: expected an indented block after function definition on line 3`, seemingly not associating the return statement as the function body when a comment precedes the `def` in this context.
## Affected Functions
All in `transformers/models/deberta_v2/modeling_deberta_v2.py`:
- `c2p_dynamic_expand` (line 104)
- `p2c_dynamic_expand` (line 110)
- `pos_dynamic_expand` (line 116)
## Workaround
Remove the comments between `@torch.jit.script` and each `def`:
```python
# Copied from transformers.models.deberta.modeling_deberta.c2p_dynamic_expand
@torch.jit.script
def c2p_dynamic_expand(c2p_pos, query_layer, relative_pos):
return c2p_pos.expand([...])
```
## Related Issues
- [pytorch/pytorch #165238](https://github.com/pytorch/pytorch/issues/165238)
- [pytorch/pytorch #65452](https://github.com/pytorch/pytorch/issues/65452)
- [pytorch/pytorch #25043](https://github.com/pytorch/pytorch/issues/25043)
- [huggingface/transformers #35443](https://github.com/huggingface/transformers/issues/35443)
- [huggingface/transformers #34190](https://github.com/huggingface/transformers/issues/34190)
- [huggingface/trl #4239](https://github.com/huggingface/trl/issues/4239)
### 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 details below)
### Reproduction
## Environment
| | |
|---|---|
| Python | 3.13.8 |
| torch | 2.10.0+cu128 |
| transformers | 5.3.0 |
| gliner | 0.2.24 |
| OS | Ubuntu 24.04.4 LTS (Noble Numbat), kernel 6.17.0-19-generic, x86_64|
## To Reproduce
```bash
python -c "from gliner import GLiNER"
```
### Expected behavior
The import should pass. | open | null | false | 4 | [
"bug"
] | [] | 2026-03-19T11:07:31Z | 2026-03-20T13:34:32Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | MNIKIEMA | 64,019,294 | MDQ6VXNlcjY0MDE5Mjk0 | User | false |
huggingface/transformers | 4,101,508,603 | I_kwDOCUB6oc70eA37 | 44,857 | https://github.com/huggingface/transformers/issues/44857 | https://api.github.com/repos/huggingface/transformers/issues/44857 | LwDetrImageLoss crashes when using float16 AMP and Cuda | ### System Info
- `transformers` version: 5.3.0.dev0
- Platform: Linux-6.6.87.2-microsoft-standard-WSL2-x86_64-with-glibc2.39
- Python version: 3.12.3
- 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?): 2.8.0+cu128 (CUDA)
- Using distributed or parallel set-up in script?: no
- Using GPU in script?: yes
- GPU type: NVIDIA RTX 500 Ada Generation Laptop GPU
### Who can help?
@yonigozlan @molbap @sbucaille
### 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)
### Reproduction
```
import torch
from transformers import LwDetrForObjectDetection
model = LwDetrForObjectDetection.from_pretrained("AnnaZhang/lwdetr_small_60e_coco")
model.train()
model.cuda()
inputs = {
"pixel_values": torch.rand((1, 3, 640, 640), device="cuda"),
"pixel_mask": torch.ones((1, 640, 640), device="cuda"),
"labels": [
{
"class_labels": torch.tensor([1], device="cuda"),
"boxes": torch.tensor([[0.5, 0.5, 0.1, 0.1]], device="cuda"),
}
],
}
with torch.autocast(device_type="cuda", dtype=torch.float16):
outputs = model(**inputs)
```
This leads to the following error:
```
".../transformers/loss/loss_lw_detr.py", line 139, in loss_labels
pos_weights[pos_ind] = t
~~~~~~~~~~~^^^^^^^^^
RuntimeError: Index put requires the source and destination dtypes match, got Half for the destination and Float for the source.
```
It seems the reason for this is that the torch.pow operations inside loss_labels() cause an upcast to float32 on the GPU, leading to a type mismatch.
### Expected behavior
Correct handling of float16 AMP on Cuda GPU, without crashing. | closed | completed | false | 3 | [
"bug"
] | [] | 2026-03-19T12:56:22Z | 2026-03-24T17:02:34Z | 2026-03-24T17:02:34Z | CONTRIBUTOR | null | 20260324T215331Z | 2026-03-24T21:53:31Z | m-matthias | 16,415,097 | MDQ6VXNlcjE2NDE1MDk3 | User | false |
huggingface/transformers | 4,102,371,828 | I_kwDOCUB6oc70hTn0 | 44,861 | https://github.com/huggingface/transformers/issues/44861 | https://api.github.com/repos/huggingface/transformers/issues/44861 | _get_tied_weight_keys crashes with AttributeError when _tied_weights_keys is a list | ### System Info
- `transformers` version: 5.3.0
- Platform: Linux
- Python version: 3.13
### Who can help?
@Cyrilvallez
### Information
- [ ] The official example scripts
- [x] My own modified scripts
### Reproduction
Full finetune of `NVIDIA-Nemotron-3-Nano-4B` crashes at checkpoint save with:
File "transformers/modeling_utils.py", line 341, in _get_tied_weight_keys
tied_weight_keys.extend([f"{name}.{k}" if name else k for k in tied.keys()])
AttributeError: 'list' object has no attribute 'keys'
The Nemotron-H model defines `_tied_weights_keys` as a list, but `_get_tied_weight_keys` assumes it is always a dict
and calls `.keys()` on it.
### Expected behavior
`save_pretrained` should handle both list and dict types for `_tied_weights_keys` without crashing. | closed | completed | false | 0 | [] | [] | 2026-03-19T15:13:12Z | 2026-03-20T09:46:56Z | 2026-03-20T09:46:56Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | gh-wf | 111,619,017 | U_kgDOBqcryQ | User | false |
huggingface/transformers | 4,102,382,384 | I_kwDOCUB6oc70hWMw | 44,863 | https://github.com/huggingface/transformers/issues/44863 | https://api.github.com/repos/huggingface/transformers/issues/44863 | NemotronH implementation can't load NemotronH checkpoints! | ### System Info
```console
- `transformers` version: 5.3.0
- Platform: macOS-15.7.3-arm64-arm-64bit
- Python version: 3.12.11
- Huggingface_hub version: 1.7.1
- Safetensors version: 0.7.0
- Accelerate version: not installed
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.10.0 (NA)
- Using distributed or parallel set-up in script?: NO
```
### Who can help?
@ArthurZucker @Cyrilvallez @liding-nv
### Information
- [x] 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)
### Reproduction
either of:
```python
tokenizer = AutoTokenizer.from_pretrained("nvidia/Nemotron-H-8B-Reasoning-128K")
model = NemotronHForCausalLM.from_pretrained("nvidia/Nemotron-H-8B-Reasoning-128K", torch_dtype=torch.bfloat16)
```
or
```python
tokenizer = AutoTokenizer.from_pretrained("nvidia/Nemotron-H-4B-Instruct-128K")
model = NemotronHForCausalLM.from_pretrained("nvidia/Nemotron-H-4B-Instruct-128K", torch_dtype=torch.bfloat16)
```
fails. The first error appears to be that these functions:
https://github.com/huggingface/transformers/blob/884333368ff329090c73bd00e57996727f301de3/src/transformers/models/nemotron_h/configuration_nemotron_h.py#L259-L269
do not correctly handle the `-` character in layer type patterns, but the saved models on the hub use them.
If I add a check to ignore the `-` character I instead see something like this:
```console
num_hidden_layers (52) is deprecated and doesn't match layers_block_type length (28). Using layers_block_type length.
Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
num_hidden_layers (52) is deprecated and doesn't match layers_block_type length (28). Using layers_block_type length.
Fetching 4 files: 100%|██████████| 4/4 [05:01<00:00, 75.32s/it]
The fast path is not available because one of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)` is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and https://github.com/Dao-AILab/causal-conv1d
Loading weights: 100%|██████████| 119/119 [00:00<00:00, 26544.82it/s]
NemotronHForCausalLM LOAD REPORT from: nvidia/Nemotron-H-8B-Reasoning-128K
Key | Status |
-------------------------------------------------+------------+-
model.layers.{4...50}.mixer.A_log | UNEXPECTED |
model.layers.{1...51}.mixer.down_proj.weight | UNEXPECTED |
model.layers.{4...50}.mixer.norm.weight | UNEXPECTED |
model.layers.{7, 18, 29, 40}.mixer.v_proj.weight | UNEXPECTED |
model.layers.{4...50}.mixer.out_proj.weight | UNEXPECTED |
model.layers.{4...50}.mixer.in_proj.weight | UNEXPECTED |
model.layers.{28...51}.norm.weight | UNEXPECTED |
model.layers.{7, 18, 29, 40}.mixer.o_proj.weight | UNEXPECTED |
model.layers.{4...50}.mixer.dt_bias | UNEXPECTED |
model.layers.{1...51}.mixer.up_proj.weight | UNEXPECTED |
model.layers.{7, 18, 29, 40}.mixer.k_proj.weight | UNEXPECTED |
model.layers.{4...50}.mixer.D | UNEXPECTED |
model.layers.{4...50}.mixer.conv1d.bias | UNEXPECTED |
model.layers.{7, 18, 29, 40}.mixer.q_proj.weight | UNEXPECTED |
model.layers.{4...50}.mixer.conv1d.weight | UNEXPECTED |
model.layers.{1...27}.mixer.out_proj.weight | MISSING |
model.layers.{1...27}.mixer.norm.weight | MISSING |
model.layers.{1...27}.mixer.A_log | MISSING |
model.layers.{1...27}.mixer.dt_bias | MISSING |
model.layers.{4, 10, 16, 22}.mixer.o_proj.weight | MISSING |
model.layers.{1...27}.mixer.conv1d.bias | MISSING |
model.layers.{4, 10, 16, 22}.mixer.k_proj.weight | MISSING |
model.layers.{1...27}.mixer.in_proj.weight | MISSING |
model.layers.{1...27}.mixer.D | MISSING |
model.layers.{1...27}.mixer.conv1d.weight | MISSING |
model.layers.{4, 10, 16, 22}.mixer.q_proj.weight | MISSING |
model.layers.{4, 10, 16, 22}.mixer.v_proj.weight | MISSING |
Notes:
- UNEXPECTED :can be ignored when loading from different task/architecture; not ok if you expect identical arch.
- MISSING :those params were newly initialized because missing from the checkpoint. Consider training on your downstream task.
```
and the model then crashes when calling `model.generate`:
```console
Traceback (most recent call last):
File "/Users/thomas/Documents/GitHub/local-lm/.venv/lib/python3.12/site-packages/transformers/tokenization_utils_base.py", line 275, in __getattr__
return self.data[item]
~~~~~~~~~^^^^^^
KeyError: 'shape'
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "/path/main.py", line 17, in <module>
outputs = model.generate(tokenized_chat)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/path/.venv/lib/python3.12/site-packages/torch/utils/_contextlib.py", line 124, in decorate_context
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/path/.venv/lib/python3.12/site-packages/transformers/generation/utils.py", line 2390, in generate
batch_size = inputs_tensor.shape[0]
^^^^^^^^^^^^^^^^^^^
File "/path/.venv/lib/python3.12/site-packages/transformers/tokenization_utils_base.py", line 277, in __getattr__
raise AttributeError
AttributeError
```
### Expected behavior
The model should output a prediction instead of crashing. | open | null | false | 4 | [
"bug"
] | [] | 2026-03-19T15:15:06Z | 2026-03-23T08:25:41Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | elfprince13 | 2,703,145 | MDQ6VXNlcjI3MDMxNDU= | User | false |
huggingface/transformers | 4,105,125,767 | I_kwDOCUB6oc70rz-H | 44,869 | https://github.com/huggingface/transformers/issues/44869 | https://api.github.com/repos/huggingface/transformers/issues/44869 | Whisper word timestamp decode crashes on trailing replacement character at end of decoded token stream | ### System Info
### System Info
- OS: macOS
- `transformers`: `5.3.0.dev0`
- Model: `openai/whisper-medium.en`
### Reproduction
I hit an `IndexError: string index out of range` in Whisper word-timestamp decoding and traced it to `src/transformers/models/whisper/tokenization_whisper.py`.
The failing code path is in `_split_tokens_on_unicode()`:
```py
decoded_full[unicode_offset + decoded.index(replacement_char)]
```
The bug happens when the decoded token stream ends with a dangling Unicode replacement character (`�`, `U+FFFD`). In that case, the computed index can equal `len(decoded_full)`, so the code reads one past the end of the string and crashes.
For the failing case I traced locally, the values were:
- `unicode_offset = 298`
- `decoded.index(replacement_char) = 0`
- `target_index = 298`
- `len(decoded_full) = 298`
So the effective access becomes:
```py
decoded_full[298]
```
but the last valid index is `297`.
The underlying ASR output for the bad chunk decoded to a long run of musical note symbols followed by a dangling final replacement character (...🎵 🎵 🎵 🎵 🎵 �). Segment-level decoding succeeded, but word-level timestamp collation crashed in `_split_tokens_on_unicode()`.
### Error
```text
IndexError: string index out of range
```
### Expected behavior
- trailing incomplete Unicode fragments at EOF should be ignored or handled safely
- Whisper word timestamp decoding should not crash with `IndexError`
### Additional context
I have a local fix prepared for this EOF bounds case and can open a PR if this approach looks reasonable.
### Who can help?
@ArthurZucker @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 details below)
### Reproduction
Full end to end reproduction would involve the original audio file (2m 14s of music with some vocals), but the underlying problem is simpler and can be reproduced by calling the `_split_tokens_on_unicode` method with data that could reasonably be outputted.
```python
from collections import defaultdict
from transformers.models.whisper.tokenization_whisper import _split_tokens_on_unicode
class DummyTokenizer:
def __init__(self):
self.responses = defaultdict(list)
def decode(self, tokens, decode_with_timestamps=False):
key = tuple(tokens)
if self.responses[key]:
return self.responses[key].pop(0)
tokenizer = DummyTokenizer()
tokenizer.responses[(1, 2)] = ["ab"] # decoded_full
tokenizer.responses[(1,)] = ["ab"] # first token decodes cleanly
tokenizer.responses[(2,)] = ["�"] # trailing replacement char at EOF
print(_split_tokens_on_unicode(tokenizer, [1, 2]))
```
Before the fix, this raises:
```text
IndexError: string index out of range
```
Because it tries to read `decoded_full[2]` when `len(decoded_full) == 2`.
### Expected behavior
Whisper word-timestamp decoding should safely ignore or stop on a trailing incomplete Unicode fragment at end-of-string, instead of crashing with `IndexError: string index out of range`. | open | null | false | 1 | [
"bug"
] | [] | 2026-03-20T01:25:49Z | 2026-03-20T13:04:50Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | chromatic-descension | 98,860,570 | U_kgDOBeR-Gg | User | false |
huggingface/transformers | 4,105,674,824 | I_kwDOCUB6oc70t6BI | 44,871 | https://github.com/huggingface/transformers/issues/44871 | https://api.github.com/repos/huggingface/transformers/issues/44871 | [Gemma-3] Inconsistent eos_token_id configuration: tokenizer has single value (1) but model.config has list [1, 106] | ### System Info
- `transformers` version: 5.3.0
- Platform: Windows-11-10.0.26100-SP0
- Python version: 3.12.11
- Huggingface_hub version: 1.7.1
- Safetensors version: 0.7.0
- Accelerate version: not installed
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.7.1+cu118 (CUDA)
- Using distributed or parallel set-up in script?: <fill in>
- Using GPU in script?: <fill in>
- GPU type: NVIDIA GeForce RTX 4060 Laptop GPU
### 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, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
MODEL = "google/gemma-3-1b-it"
tokenizer = AutoTokenizer.from_pretrained(MODEL)
model = AutoModelForCausalLM.from_pretrained(MODEL)
print(tokenizer.eos_token_id) # 1
print(model.config.eos_token_id) # [1, 106]
```
### Expected behavior
## Expected Behavior
Tokenizer and Model Config should be consistent on eos_token_id so that the following generation loop can work correctly:
```python
for _ in range(max_new_tokens):
...
if new_token_id.item() == tokenizer.eos_token_id:
break # should break the loop
``` | closed | completed | false | 1 | [
"bug"
] | [] | 2026-03-20T04:29:57Z | 2026-03-21T01:40:09Z | 2026-03-21T01:40:09Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | IvanFan-Van | 98,149,954 | U_kgDOBdmmQg | User | false |
huggingface/transformers | 4,106,938,437 | I_kwDOCUB6oc70yuhF | 44,877 | https://github.com/huggingface/transformers/issues/44877 | https://api.github.com/repos/huggingface/transformers/issues/44877 | Strict config prevents loading `granite_speech` config | ### System Info
- `transformers` version: 5.3.0.dev0
- Platform: Windows-10-10.0.26200-SP0
- Python version: 3.11.13
- Huggingface_hub version: 1.7.1
- Safetensors version: 0.6.2
- Accelerate version: 1.11.0
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.9.0+cu126 (CUDA)
- Using distributed or parallel set-up in script?: No
- Using GPU in script?: No
- GPU type: NVIDIA GeForce RTX 3090
### 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)
### Reproduction
```python
from transformers import AutoConfig
model_name = "ibm-granite/granite-4.0-1b-speech"
config = AutoConfig.from_pretrained(model_name)
```
```
Traceback (most recent call last):
File "[sic]/Lib/site-packages/huggingface_hub/dataclasses.py", line 148, in __strict_setattr__
validator(value)
File "[sic]/Lib/site-packages/huggingface_hub/dataclasses.py", line 607, in validator
type_validator(field.name, value, field.type)
File "[sic]/Lib/site-packages/huggingface_hub/dataclasses.py", line 474, in type_validator
_validate_simple_type(name, value, expected_type)
File "[sic]/Lib/site-packages/huggingface_hub/dataclasses.py", line 597, in _validate_simple_type
raise TypeError(
TypeError: Field 'embedding_multiplier' expected float, got int (value: 12)
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "[sic]/demo_granite_4_1b_speech.py", line 4, in <module>
config = AutoConfig.from_pretrained(model_name)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "[sic]/src/transformers/models/auto/configuration_auto.py", line 1486, in from_pretrained
return config_class.from_dict(config_dict, **unused_kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "[sic]/src/transformers/configuration_utils.py", line 757, in from_dict
config = cls(**config_dict)
^^^^^^^^^^^^^^^^^^
File "[sic]/Lib/site-packages/huggingface_hub/dataclasses.py", line 279, in init_with_validate
initial_init(self, *args, **kwargs) # type: ignore [call-arg]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "[sic]/Lib/site-packages/huggingface_hub/dataclasses.py", line 194, in __init__
self.__post_init__(**additional_kwargs)
File "[sic]/src/transformers/models/granite_speech/configuration_granite_speech.py", line 122, in __post_init__
self.text_config = CONFIG_MAPPING[self.text_config["model_type"]](**self.text_config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "[sic]/Lib/site-packages/huggingface_hub/dataclasses.py", line 279, in init_with_validate
initial_init(self, *args, **kwargs) # type: ignore [call-arg]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "[sic]/Lib/site-packages/huggingface_hub/dataclasses.py", line 179, in __init__
setattr(self, f.name, standard_kwargs[f.name])
File "[sic]/Lib/site-packages/huggingface_hub/dataclasses.py", line 150, in __strict_setattr__
raise StrictDataclassFieldValidationError(field=name, cause=e) from e
huggingface_hub.errors.StrictDataclassFieldValidationError: Validation error for field 'embedding_multiplier':
TypeError: Field 'embedding_multiplier' expected float, got int (value: 12)
```
### Expected behavior
For the config to load as normal.
- Tom Aarsen | open | null | false | 4 | [
"bug"
] | [] | 2026-03-20T09:57:02Z | 2026-03-20T10:47:03Z | null | MEMBER | null | 20260324T120231Z | 2026-03-24T12:02:31Z | tomaarsen | 37,621,491 | MDQ6VXNlcjM3NjIxNDkx | User | false |
huggingface/transformers | 4,109,855,475 | I_kwDOCUB6oc7092rz | 44,898 | https://github.com/huggingface/transformers/issues/44898 | https://api.github.com/repos/huggingface/transformers/issues/44898 | [BUG] Perceiver image classification (non-default res) fails even with interpolate_pos_encoding=True | ### 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: `not installed`
* PyTorch version (accelerator?): `2.9.1+cu128 (CUDA)`
* GPU type: `NVIDIA L4`
* NVIDIA driver version: `550.90.07`
* CUDA version: `12.4`
### Information
- [x] The official example scripts
- [ ] My own modified scripts
### Tasks
- [x] An officially supported task in the `examples` folder (such as GLUE/SQuAD, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
```python
import torch
from transformers import PerceiverForImageClassificationLearned, PerceiverImageProcessorPil
from PIL import Image
import requests
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
image = Image.open(requests.get(url, stream=True).raw)
image_processor = PerceiverImageProcessorPil(size={"height": 384, "width": 384})
model = PerceiverForImageClassificationLearned.from_pretrained("deepmind/vision-perceiver-learned")
model.eval()
inputs = image_processor(image, return_tensors="pt").pixel_values
try:
with torch.no_grad():
outputs = model(inputs=inputs, interpolate_pos_encoding=True)
print("Logits shape:", outputs.logits.shape)
predicted_class = outputs.logits.argmax(-1).item()
print("Predicted class:", predicted_class)
except Exception as e:
print(e)
```
→ Trying to run image classification on a 384×384 image (pretrained default is 224×224) and even after setting `interpolate_pos_encoding=True` expecting the model to handle the resolution difference, the model crashes with a `RuntimeError`.
→ From the screenshot, 384×384 = 147456 and 224×224 = 50176 so it was never actually resized (see the reproduction output).
**Current Repro Output:**
<img width="500" height="500" alt="Image" src="https://github.com/user-attachments/assets/3f1ac00d-5f36-4d3b-be2a-21f46accd0bb" />
### Expected behavior
→ Inference should complete successfully (torch.Size([1, 1000])) when `interpolate_pos_encoding=True` is passed with non-native input res. | open | null | false | 0 | [
"bug"
] | [] | 2026-03-20T19:58:09Z | 2026-03-20T19:58:24Z | null | CONTRIBUTOR | null | 20260324T120231Z | 2026-03-24T12:02:31Z | harshaljanjani | 75,426,551 | MDQ6VXNlcjc1NDI2NTUx | User | false |
huggingface/transformers | 4,111,564,283 | I_kwDOCUB6oc71EX37 | 44,906 | https://github.com/huggingface/transformers/issues/44906 | https://api.github.com/repos/huggingface/transformers/issues/44906 | Remove unnecessary `expand_as` in `get_placeholder_mask` across VLMs | ### Feature request
## Problem
The `get_placeholder_mask` function (and equivalent inline patterns) across ~70 multimodal model files expands a boolean placeholder mask from shape `(B, S, 1)` to `(B, S, H)` via `.expand_as(inputs_embeds)` before passing it to `masked_scatter`. This expansion is unnecessary because `masked_scatter` natively supports broadcasting.
For example, in `modeling_llava.py`:
```python
special_image_mask = special_image_mask.unsqueeze(-1).expand_as(inputs_embeds).to(inputs_embeds.device)
```
The validation check also uses data-dependent boolean indexing:
```python
inputs_embeds[special_image_mask].numel() == image_features.numel()
```
### Motivation
## Motivation
Mainly, Memory. While `expand_as` itself returns a stride-0 view (no copy), the subsequent `.to(device)` call materializes the full `(B, S, H)` boolean tensor when a device transfer is needed. Practically, I assume that the need for actual device transfer is quite rare; however, I believe this is still a safer implementation.
### Your contribution
## Proposed fix
1. Remove `.expand_as(inputs_embeds)`, keeping the mask as `(B, S, 1)`:
```python
special_image_mask = special_image_mask.unsqueeze(-1).to(inputs_embeds.device)
```
2. Replace the validation with equivalent arithmetic:
```python
n_image_tokens * inputs_embeds.shape[-1] == image_features.numel()
```
## Correctness
`masked_scatter`, `torch.where`, and element-wise `*` all support broadcasting a `(B, S, 1)` mask to `(B, S, H)`. Verified with:
```python
import torch
B, S, H = 2, 10, 16
inputs_embeds = torch.randn(B, S, H)
features = torch.randn(3, H)
mask_2d = torch.zeros(B, S, dtype=torch.bool)
mask_2d[0, 2], mask_2d[0, 5], mask_2d[1, 3] = True, True, True
mask_old = mask_2d.unsqueeze(-1).expand_as(inputs_embeds)
mask_new = mask_2d.unsqueeze(-1)
result_old = inputs_embeds.clone().masked_scatter(mask_old, features)
result_new = inputs_embeds.clone().masked_scatter(mask_new, features)
assert torch.equal(result_old, result_new) # ✓ identical
```
Affects `get_placeholder_mask` and equivalent inline patterns in ~70 files across all VLM models (llava, qwen2_vl, paligemma, gemma3n, chameleon, video_llava, etc.), plus `tensor_parallel.py` and `ovis2`.
I have a PR ready: https://github.com/syncdoth/transformers/tree/remove-expand-as-placeholder-mask | open | null | false | 0 | [
"Feature request"
] | [] | 2026-03-21T06:05:36Z | 2026-03-21T06:05:36Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | syncdoth | 45,599,998 | MDQ6VXNlcjQ1NTk5OTk4 | User | false |
huggingface/transformers | 4,111,661,658 | I_kwDOCUB6oc71Evpa | 44,908 | https://github.com/huggingface/transformers/issues/44908 | https://api.github.com/repos/huggingface/transformers/issues/44908 | inverse_sqrt scheduler ignores lr_scheduler_kwargs (timescale not passed) | ### System Info
Incomplete arguments passed for schedulers where name is explicitly checked.
https://github.com/huggingface/transformers/blob/v5.3.0/src/transformers/optimization.py#L664
### 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, ...)
- [ ] My own task or dataset (give details below)
### Reproduction
Trigger a learning job with "inverse_sqrt" and set timesteps = 'random', it still works because the arguments are never passed.
### Expected behavior
timesteps should take effect. | closed | completed | false | 1 | [
"bug"
] | [] | 2026-03-21T06:53:50Z | 2026-03-24T13:06:18Z | 2026-03-24T13:06:18Z | NONE | null | 20260324T143453Z | 2026-03-24T14:34:53Z | magarwal0205 | 6,791,598 | MDQ6VXNlcjY3OTE1OTg= | User | false |
huggingface/transformers | 4,112,866,232 | I_kwDOCUB6oc71JVu4 | 44,910 | https://github.com/huggingface/transformers/issues/44910 | https://api.github.com/repos/huggingface/transformers/issues/44910 | [Bug] Flash Attention crashes with illegal memory access on Qwen3.5 due to 3D position_ids being misinterpreted as packed sequence | ### System Info
# [Bug] Flash Attention crashes with `illegal memory access` on Qwen3.5 due to 3D `position_ids` being misinterpreted as packed sequence
We fixed it in https://github.com/ouroborosscr/transformers/tree/fix/qwen35-flash-attn-3d-position-ids
## Description
When using `attn_implementation="flash_attention_2"` with Qwen3.5 models, all forward passes crash with `CUDA error: an illegal memory access was encountered`. This affects both training and inference.
**Root cause**: Qwen3.5 uses a hybrid architecture (GatedDeltaNet linear attention + standard attention) and passes **3D `position_ids`** with shape `[3, batch_size, seq_len]` (for multi-dimensional rotary embedding). The function `_is_packed_sequence()` in `modeling_flash_attention_utils.py` misinterprets this 3D tensor as a packed sequence indicator, causing `cu_seqlens` to be constructed with 3× the actual token count. Flash Attention then reads beyond the q/k/v tensor boundaries, resulting in an illegal memory access.
## Reproduction
```python
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen3.5-9B", # or any Qwen3.5 variant
quantization_config=BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
),
torch_dtype=torch.bfloat16,
device_map={"": 0},
attn_implementation="flash_attention_2",
)
# This crashes immediately
input_ids = torch.randint(100, 5000, (1, 256), device="cuda")
with torch.no_grad():
out = model(input_ids=input_ids, use_cache=False)
```
**Error**:
```
torch.AcceleratorError: CUDA error: an illegal memory access was encountered
```
**Traceback** (abbreviated):
```
File "transformers/modeling_flash_attention_utils.py", line 692, in _flash_attention_forward
out = flash_varlen_fn(
File "flash_attn/flash_attn_interface.py", line 1443, in flash_attn_varlen_func
return FlashAttnVarlenFunc.apply(
File "flash_attn/flash_attn_interface.py", line 165, in _flash_attn_varlen_forward
out, softmax_lse, S_dmask, rng_state = flash_attn_gpu.varlen_fwd(
torch.AcceleratorError: CUDA error: an illegal memory access was encountered
```
## Root Cause Analysis
### Qwen3.5 hybrid architecture
Qwen3.5 uses a **mixed attention architecture**: 24 layers of `Qwen3_5GatedDeltaNet` (linear attention) and 8 layers of `Qwen3_5Attention` (standard attention, at layers 3, 7, 11, 15, 19, 23, 27, 31). Only the standard attention layers use flash attention.
Qwen3.5 passes **3D `position_ids`** with shape `[3, batch_size, seq_len]` for its multi-dimensional rotary embedding (3 sets of position indices).
### The bug
In `modeling_flash_attention_utils.py`, the function `_is_packed_sequence()` (line 444) does not handle tensors with more than 2 dimensions:
```python
def _is_packed_sequence(position_ids, batch_size):
if position_ids is None:
return False
increasing_position_sequences = (
torch.arange(position_ids.shape[1], device=position_ids.device) + position_ids.min()
)
return batch_size == 1 and (increasing_position_sequences - position_ids).abs().sum().bool()
```
When `position_ids` has shape `[3, 1, 256]`:
- `position_ids.shape[1]` returns `1` (not `256` as expected for a 2D `[batch, seq_len]` tensor)
- The function returns `True`, misidentifying this as a packed sequence
This triggers the packed-sequence code path at line 677:
```python
elif is_fa_with_varlen_kwargs or is_fa_with_position_ids:
if cu_seq_lens_q is None or cu_seq_lens_k is None:
q, k, v, (cu_seq_lens_q, cu_seq_lens_k), (max_length_q, max_length_k) = _prepare_from_posids(
query_states, key_states, value_states, position_ids
)
```
Inside `prepare_fa_kwargs_from_position_ids()` (line 362):
```python
position_ids = position_ids.reshape(-1) # [3, 1, 256] → [768]
indices_q = (position_ids == 0).nonzero().view(-1) # Finds 3 zero positions
```
This constructs `cu_seqlens = [0, 256, 512, 768]`, claiming 3 sequences with 768 total tokens. But the actual q/k/v tensors only contain 256 tokens. Flash Attention reads up to index 768, causing the illegal memory access.
### Intercepted parameters confirming the mismatch
```
🔍 varlen_fwd parameters:
q: torch.Size([256, 16, 256]) ← 256 tokens
cu_seqlens_q: tensor([0, 256, 512, 768]) ← claims 768 tokens
q total=256 vs cu_seqlens_q[-1]=768 ← MISMATCH → crash
```
## Fix
Add a dimensionality check in `_is_packed_sequence()` to reject tensors with more than 2 dimensions, since packed sequences always use 2D position_ids `[batch, seq_len]`:
```python
def _is_packed_sequence(position_ids, batch_size):
if position_ids is None:
return False
if position_ids.dim() > 2:
return False
increasing_position_sequences = (
torch.arange(position_ids.shape[1], device=position_ids.device) + position_ids.min()
)
return batch_size == 1 and (increasing_position_sequences - position_ids).abs().sum().bool()
```
This fix has been validated: all 8 standard attention layers in Qwen3.5-9B pass flash attention forward successfully after applying the patch.
## Environment
- **Model**: Qwen3.5-9B (hybrid GatedDeltaNet + standard attention)
- **GPU**: NVIDIA A100-SXM4-80GB
- **PyTorch**: 2.9.0 / 2.10.0 (both affected)
- **Transformers**: 5.3.0
- **flash-attn**: 2.8.3
- **CUDA**: 12.8
## Impact
- Affects **all Qwen3.5 variants** (and potentially any future model using >2D position_ids)
- Blocks both training and inference when using `flash_attention_2`
- No workaround other than falling back to `sdpa` or `eager` attention implementations
### Who can help?
@vasqu @ArthurZucker (attention)
### 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)
### Reproduction
GRPO reinforcement learning training with Qwen3.5-9B using TRL GRPOTrainer
When using `attn_implementation="flash_attention_2"` with Qwen3.5, all forward passes crash with CUDA illegal memory access. Minimal reproduction:
```python
import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen3.5-9B",
quantization_config=BitsAndBytesConfig(
load_in_4bit=True,
bnb_4bit_compute_dtype=torch.bfloat16,
),
torch_dtype=torch.bfloat16,
device_map={"": 0},
attn_implementation="flash_attention_2",
)
input_ids = torch.randint(100, 5000, (1, 256), device="cuda")
with torch.no_grad():
out = model(input_ids=input_ids, use_cache=False) # crashes here
```
**Error:**
```
torch.AcceleratorError: CUDA error: an illegal memory access was encountered
```
**Root cause:**
Qwen3.5 is a hybrid architecture (24 GatedDeltaNet layers + 8 standard attention layers). It uses 3D `position_ids` with shape `[3, batch_size, seq_len]` for multi-dimensional rotary embedding.
`_is_packed_sequence()` in `modeling_flash_attention_utils.py` (line 444) does not handle >2D tensors:
```python
def _is_packed_sequence(position_ids, batch_size):
if position_ids is None:
return False
increasing_position_sequences = (
torch.arange(position_ids.shape[1], device=position_ids.device) + position_ids.min()
)
return batch_size == 1 and (increasing_position_sequences - position_ids).abs().sum().bool()
```
When `position_ids` has shape `[3, 1, 256]`, `position_ids.shape[1]` returns `1` instead of the sequence length, and the function returns `True`, misidentifying this as a packed sequence.
This triggers `prepare_fa_kwargs_from_position_ids()` which does:
```python
position_ids = position_ids.reshape(-1) # [3, 1, 256] → [768]
indices_q = (position_ids == 0).nonzero() # finds 3 zero positions
# constructs cu_seqlens = [0, 256, 512, 768] — claims 768 tokens
```
But q/k/v only contain 256 tokens. Flash attention reads up to index 768, causing the illegal memory access.
**Intercepted evidence:**
```
q: torch.Size([256, 16, 256]) ← 256 tokens
cu_seqlens_q: tensor([0, 256, 512, 768]) ← claims 768 tokens
q total=256 vs cu_seqlens_q[-1]=768 ← MISMATCH → crash
```
**Environment:**
- Model: Qwen3.5-9B
- GPU: NVIDIA A100-SXM4-80GB
- PyTorch: 2.9.0 and 2.10.0 (both affected)
- transformers: 5.3.0
- flash-attn: 2.8.3
- CUDA: 12.8
**Fix:** Add `if position_ids.dim() > 2: return False` in `_is_packed_sequence()`. Packed sequences always use 2D `[batch, seq_len]` position_ids.
### Expected behavior
Model forward pass with `attn_implementation="flash_attention_2"` should complete successfully without CUDA errors. After the fix (adding dimensionality check for >2D position_ids), all 8 standard attention layers in Qwen3.5-9B pass flash attention forward correctly. | closed | completed | false | 5 | [
"bug"
] | [] | 2026-03-21T15:38:54Z | 2026-03-24T14:36:36Z | 2026-03-24T14:36:36Z | NONE | null | 20260324T145640Z | 2026-03-24T14:56:40Z | ouroborosscr | 75,082,702 | MDQ6VXNlcjc1MDgyNzAy | User | false |
huggingface/transformers | 4,113,147,717 | I_kwDOCUB6oc71KadF | 44,912 | https://github.com/huggingface/transformers/issues/44912 | https://api.github.com/repos/huggingface/transformers/issues/44912 | git-oss-20b will not properly load with MXFP4 quantization and falls back to bf16 | ### System Info
Hi, probably this is related to
https://github.com/huggingface/transformers/issues/42723
I get:
MXFP4 quantization requires Triton and kernels installed: CUDA requires Triton >= 3.4.0, XPU requires Triton >= 3.5.0, we will default to dequantizing the model to bf16
When executing the following code:
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
token=HF_TOKEN,
#model_kwargs={"dtype": torch.bfloat16},
device_map="auto",
)
$ hf env
Copy-and-paste the text below in your GitHub issue.
- huggingface_hub version: 1.7.2
- Platform: Linux-6.8.0-106-generic-x86_64-with-glibc2.39
- Python version: 3.12.3
- Running in iPython ?: No
- Running in notebook ?: No
- Running in Google Colab ?: No
- Running in Google Colab Enterprise ?: No
- Token path ?: /scratch/work/ml_ki/llama/huggingface/token
- Has saved token ?: True
- Who am I ?: jottbe
- Configured git credential helpers:
- Installation method: unknown
- httpx: 0.28.1
- hf_xet: 1.4.2
- gradio: N/A
- tensorboard: N/A
- ENDPOINT: https://huggingface.co
- HF_HUB_CACHE: /scratch/work/ml_ki/llama/huggingface/hub
- HF_ASSETS_CACHE: /scratch/work/ml_ki/llama/huggingface/assets
- HF_TOKEN_PATH: /scratch/work/ml_ki/llama/huggingface/token
- HF_STORED_TOKENS_PATH: /scratch/work/ml_ki/llama/huggingface/stored_tokens
- HF_HUB_OFFLINE: False
- HF_HUB_DISABLE_TELEMETRY: False
- HF_HUB_DISABLE_PROGRESS_BARS: None
- HF_HUB_DISABLE_SYMLINKS_WARNING: False
- HF_HUB_DISABLE_EXPERIMENTAL_WARNING: False
- HF_HUB_DISABLE_IMPLICIT_TOKEN: False
- HF_HUB_DISABLE_XET: False
- HF_HUB_ETAG_TIMEOUT: 10
- HF_HUB_DOWNLOAD_TIMEOUT: 10
- HF_XET_HIGH_PERFORMANCE: False
Package versions:
transformers 5.3.0
triton 3.6.0
nvidia-cublas-cu12 12.8.4.1
nvidia-cuda-cupti-cu12 12.8.90
nvidia-cuda-nvrtc-cu12 12.8.93
nvidia-cuda-runtime-cu12 12.8.90
nvidia-cudnn-cu12 9.10.2.21
nvidia-cufft-cu12 11.3.3.83
nvidia-cufile-cu12 1.13.1.3
nvidia-curand-cu12 10.3.9.90
nvidia-cusolver-cu12 11.7.3.90
nvidia-cusparse-cu12 12.5.8.93
nvidia-cusparselt-cu12 0.7.1
nvidia-nccl-cu12 2.27.5
nvidia-nvjitlink-cu12 12.8.93
nvidia-nvshmem-cu12 3.4.5
nvidia-nvtx-cu12 12.8.90
torch 2.10.0
torchaudio 2.10.0
torchvision 0.25.0
$ nvcc --version
nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2023 NVIDIA Corporation
Built on Fri_Jan__6_16:45:21_PST_2023
Cuda compilation tools, release 12.0, V12.0.140
Build cuda_12.0.r12.0/compiler.32267302_0
### Who can help?
@ArthurZucker @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)
### Reproduction
pipeline = transformers.pipeline(
"text-generation",
model=model_id,
token=HF_TOKEN,
#model_kwargs={"dtype": torch.bfloat16},
device_map="auto",
)
### Expected behavior
Should load with 4bit quantization. | closed | completed | false | 5 | [
"bug"
] | [] | 2026-03-21T17:38:02Z | 2026-03-24T15:33:17Z | 2026-03-24T15:33:17Z | NONE | null | 20260324T215331Z | 2026-03-24T21:53:31Z | jottbecr | 126,362,918 | U_kgDOB4glJg | User | false |
huggingface/transformers | 4,113,212,978 | I_kwDOCUB6oc71KqYy | 44,913 | https://github.com/huggingface/transformers/issues/44913 | https://api.github.com/repos/huggingface/transformers/issues/44913 | In GPTNeoXConfig, rotary_pct silently reverts to default on reload | ### System Info
- `transformers` version: 5.3.0
- Platform: Linux-6.17.0-19-generic-x86_64-with-glibc2.39
- Python version: 3.12.4
- Huggingface_hub version: 1.7.2
- Safetensors version: 0.4.5
- Accelerate version: not installed
- Accelerate config: not found
- DeepSpeed version: not installed
- PyTorch version (accelerator?): 2.5.1 (NA)
- Using distributed or parallel set-up in script?: No
When creating a `GPTNeoXConfig` with non-default `rotary_pct`, value is lost after `save_pretrained` / `from_pretrained`.
## Cause
https://github.com/huggingface/transformers/blob/3a3b59cb1a7c0238c8d1072e35d3879c5faff48e/src/transformers/models/gpt_neox/configuration_gpt_neox.py#L98
`save_pretrained` writes `partial_rotary_factor` inside `rope_parameters` but does not persist `rotary_pct` as a top-level key. On reload, `rotary_pct` is absent from kwargs, so this line unconditionally overwrites the correct value with `0.25`.
## Fix
```python
rotary_pct = kwargs.pop("rotary_pct", None)
if rotary_pct is not None:
self.rope_parameters["partial_rotary_factor"] = rotary_pct
else:
self.rope_parameters.setdefault("partial_rotary_factor", 0.25)
```
Verified locally after applying this, the value survives the round-trip.
Models using the default `rotary_pct=0.25` (gpt-neox-20b, Pythia, etc.) are unaffected since the overwrite produces the same value.
### Who can help?
@ArthurZucker @Cyrilvallez
### 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)
### Reproduction
```python
from transformers import GPTNeoXConfig
config = GPTNeoXConfig(rotary_pct=1.0)
print(config.rope_parameters["partial_rotary_factor"]) # 1.0
config.save_pretrained("/tmp/test")
config2 = GPTNeoXConfig.from_pretrained("/tmp/test")
print(config2.rope_parameters["partial_rotary_factor"]) # 0.25
```
### Expected behavior
`partial_rotary_factor` value should be retained | open | null | false | 1 | [
"bug"
] | [] | 2026-03-21T17:58:32Z | 2026-03-23T10:09:34Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | ratishsp | 3,006,607 | MDQ6VXNlcjMwMDY2MDc= | User | false |
huggingface/transformers | 4,114,204,538 | I_kwDOCUB6oc71Ocd6 | 44,918 | https://github.com/huggingface/transformers/issues/44918 | https://api.github.com/repos/huggingface/transformers/issues/44918 | Unpacking Qwen3.5 input embeddings fails with trl SFT trainer | ### System Info
My environment:
```
datasets==4.6.1
faiss-cpu==1.13.2
numpy==2.4.2
pyserini==1.5.0
sentence-transformers==5.2.3
torch==2.10.0
torchvision==0.25.0
tqdm==4.67.3
trl==0.29.1
wandb==0.25.1
```
### 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)
### Reproduction
This is the code I am using:
```py
config = SFTConfig(
output_dir=checkpoints_path,
max_steps=num_steps,
dataset_text_field="text",
max_length=2048,
logging_strategy="steps",
logging_steps=max(1, checkpoint_frequency // 5),
disable_tqdm=False,
save_strategy="steps",
save_steps=checkpoint_frequency,
eval_strategy="steps",
eval_steps=checkpoint_frequency,
load_best_model_at_end=True,
metric_for_best_model="eval_loss",
greater_is_better=False,
report_to="wandb",
run_name=run_name,
)
trainer = SFTTrainer(
model=lm.model,
args=config,
train_dataset=train_ds,
eval_dataset=eval_ds,
)
trainer.train()
trainer.save_model()
```
### Expected behavior
Successful SFT. Instead I get the following error with a simple qwen3.5 SFT setup. Running inference only worked fine for me
```
Traceback (most recent call last):
File "/workspace/writeable/open-ended-csp/tmp.py", line 41, in <module>
train_sft(
File "/workspace/writeable/open-ended-csp/utils/sft_trainer.py", line 52, in train_sft
trainer.train()
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/transformers/trainer.py", line 1412, in train
return inner_training_loop(
^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/transformers/trainer.py", line 1742, in _inner_training_loop
tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/trl/trainer/sft_trainer.py", line 1338, in training_step
return super().training_step(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/transformers/trainer.py", line 1951, in training_step
loss = self.compute_loss(model, inputs, num_items_in_batch=num_items_in_batch)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/trl/trainer/sft_trainer.py", line 1234, in compute_loss
(loss, outputs) = super().compute_loss(
^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/transformers/trainer.py", line 2022, in compute_loss
outputs = model(**inputs)
^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/accelerate/utils/operations.py", line 823, in forward
return model_forward(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/accelerate/utils/operations.py", line 811, in __call__
return convert_to_fp32(self.model_forward(*args, **kwargs))
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/torch/amp/autocast_mode.py", line 44, in decorate_autocast
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/transformers/utils/generic.py", line 841, in wrapper
output = func(self, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/transformers/models/qwen3_5/modeling_qwen3_5.py", line 1938, in forward
outputs = self.model(
^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1776, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1787, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/transformers/utils/generic.py", line 841, in wrapper
output = func(self, *args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/transformers/models/qwen3_5/modeling_qwen3_5.py", line 1685, in forward
position_ids = self.compute_3d_position_ids(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/workspace/writeable/open-ended-csp/.venv/lib/python3.12/site-packages/transformers/models/qwen3_5/modeling_qwen3_5.py", line 1619, in compute_3d_position_ids
batch_size, seq_length, _ = inputs_embeds.shape
^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: too many values to unpack (expected 3)
wandb:
wandb: 🚀 View run test-qwen3.5 at: https://wandb.ai/jj-uoe/open-ended-csp/runs/1sercncq
wandb: Find logs at: wandb/run-20260321_232655-1sercncq/logs
``` | closed | completed | false | 3 | [
"bug"
] | [] | 2026-03-21T23:43:15Z | 2026-03-23T15:20:11Z | 2026-03-23T15:20:11Z | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | JakobJBauer | 53,357,351 | MDQ6VXNlcjUzMzU3MzUx | User | false |
huggingface/transformers | 4,116,497,241 | I_kwDOCUB6oc71XMNZ | 44,928 | https://github.com/huggingface/transformers/issues/44928 | https://api.github.com/repos/huggingface/transformers/issues/44928 | [Bug] Catastrophic gradient explosion (NaN) in RLHF with Qwen3.5 due to 3D position_ids forcing SDPA Math fallback and BF16 collapse | ### System Info
Copy-and-paste the text below in your GitHub issue and FILL OUT the two last points.
- `transformers` version: 5.3.0
- Platform: Linux-6.8.0-40-generic-x86_64-with-glibc2.35
- Python version: 3.11.15
- Huggingface_hub version: 1.7.1
- Safetensors version: 0.7.0
- Accelerate version: 1.13.0
- Accelerate config: not found
- DeepSpeed version: 0.18.8
- PyTorch version (accelerator?): 2.10.0+cu128 (CUDA)
- Using distributed or parallel set-up in script?: <fill in>
- Using GPU in script?: <fill in>
- GPU type: NVIDIA A100-SXM4-80GB
### Who can help?
@ArthurZucker @Cyrilvallez
### 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)
### Reproduction
We have fully isolated the issue and provided a reproducible repository here: 👉 https://github.com/ouroborosscr/Report-the-gradient-explosion-of-qwen3.5
1.Verify Dapo degradation
```python
CUDA_LAUNCH_BLOCKING=1 CUDA_HOME=/usr/local/cuda-12.8 LD_PRELOAD=/opt/anaconda3/envs/scr_train2/lib/libstdc++.so.6 CUDA_VISIBLE_DEVICES=2 torchrun --nproc_per_node=1 audit_qwen_hf.py
```
Output:
```
============================================================
🔍 [源码级审计] Qwen3.5 (SDPA 模式) 到底传了什么给底层？
============================================================
➡️ 模拟 DAPO 训练输入，序列长度: 8192
🚨 [劫持成功] Qwen3.5 正在调用 PyTorch SDPA!
Query 形状: torch.Size([1, 8, 8192, 128])
⚠️ Qwen 传来的 Mask 形状: torch.Size([1, 1, 8192, 8192])
⚠️ Mask 的数据类型: torch.bool
💀 仅这个 Mask 矩阵就会占据显存: 0.125000 GB
is_causal 参数: False
```
2.Verify gradient explosion
```python
CUDA_LAUNCH_BLOCKING=1 CUDA_HOME=/usr/local/cuda-12.8 LD_PRELOAD=/opt/anaconda3/envs/scr_train2/lib/libstdc++.so.6 CUDA_VISIBLE_DEVICES=2 torchrun --nproc_per_node=1 train_dapo_3_debug.py --use_lora --use_4bit
```
Output:
```
……
💥 [爆点定位] 梯度在经过 【model.layers.30.linear_attn.norm】 的反向计算后瞬间爆炸！
传出梯度最大值: 26240.0
⚠️ [数值爆炸预警] 层: model.layers.27.self_attn.v_proj | 梯度 Max: 17280.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.27.self_attn.k_norm | 梯度 Max: 135168.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.27.self_attn.k_proj | 梯度 Max: 111616.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.27.self_attn.q_norm | 梯度 Max: 126976.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.27.self_attn.q_proj | 梯度 Max: 97280.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.23.self_attn | 梯度 Max: 67072.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.23.self_attn.o_proj | 梯度 Max: 67072.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.23.self_attn.v_proj | 梯度 Max: 13172736.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.23.self_attn.k_norm | 梯度 Max: 62390272.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.23.self_attn.k_proj | 梯度 Max: 45613056.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.23.self_attn.q_norm | 梯度 Max: 297795584.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.23.self_attn.q_proj | 梯度 Max: 242221056.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.19.self_attn | 梯度 Max: 132120576.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.19.self_attn.o_proj | 梯度 Max: 132120576.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.19.self_attn.v_proj | 梯度 Max: 48103633715200.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.19.self_attn.k_norm | 梯度 Max: 68444598829056.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.19.self_attn.k_proj | 梯度 Max: 54150947667968.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.19.self_attn.q_norm | 梯度 Max: 141836999983104.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.19.self_attn.q_proj | 梯度 Max: 204509162766336.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.15.self_attn | 梯度 Max: 131391639519232.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.15.self_attn.o_proj | 梯度 Max: 131391639519232.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.15.self_attn.v_proj | 梯度 Max: 1549526502191602335744.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.15.self_attn.k_norm | 梯度 Max: 1752440687002407403520.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.15.self_attn.k_proj | 梯度 Max: 2822351843277561397248.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.15.self_attn.q_norm | 梯度 Max: 2600990914393046777856.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.15.self_attn.q_proj | 梯度 Max: 5902958103587056517120.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.11.self_attn | 梯度 Max: 2951479051793528258560.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.11.self_attn.o_proj | 梯度 Max: 2951479051793528258560.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.11.self_attn.v_proj | 梯度 Max: 4584246707978673830485819392.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.11.self_attn.k_norm | 梯度 Max: 7969239002899635519663112192.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.11.self_attn.k_proj | 梯度 Max: 14932651723879899565970685952.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.11.self_attn.q_norm | 梯度 Max: 40852021296417549071671099392.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.11.self_attn.q_proj | 梯度 Max: 41470991316060239209120661504.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.7.self_attn | 梯度 Max: 52612451669628661683212779520.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.7.self_attn.o_proj | 梯度 Max: 52612451669628661683212779520.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.7.self_attn.v_proj | 梯度 Max: 7382797095729208034316799468109824.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.7.self_attn.k_norm | 梯度 Max: 29206669829258405410484041851863040.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.7.self_attn.k_proj | 梯度 Max: 50949412924372996104955495230472192.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.7.self_attn.q_norm | 梯度 Max: 55168154121932543553136523497963520.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.7.self_attn.q_proj | 梯度 Max: 113581493780449354374104607201689600.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.3.self_attn | 梯度 Max: 138893940965806639063190776806637568.00 | Dtype: torch.bfloat16
⚠️ [数值爆炸预警] 层: model.layers.3.self_attn.o_proj | 梯度 Max: 138893940965806639063190776806637568.00 | Dtype: torch.bfloat16
……
```
3.Fix
```
CUDA_LAUNCH_BLOCKING=1 CUDA_HOME=/usr/local/cuda-12.8 LD_PRELOAD=/opt/anaconda3/envs/scr_train2/lib/libstdc++.so.6 CUDA_VISIBLE_DEVICES=2 torchrun --nproc_per_node=1 train_dapo_3_debug.py --use_lora --use_4bit --use_flash_attn
```
Output:
```
……
[🔍 22:47:55] ✅ Step 1 梯度全部正常 (最大 =0.003845 @ base_model.model.model.layers.3.self_attn.v_proj.lora_B.default.weight)
……
```
### Expected behavior
1. **Avoid Silent Fallbacks to the Math Backend:** When training Qwen3.5 models (or any Qwen2 architecture handling 3D `position_ids`/mRoPE), the `transformers` implementation explicitly materializes a massive 4D Dense Mask (`[Batch, 1, SeqLen, SeqLen]`) and sets `is_causal=False`. This design explicitly violates PyTorch SDPA’s fused kernel constraints (`if (attn_mask.has_value()) { return false; }`), silently forcing a downgrade to the `Math` backend. **Expected behavior:** The implementation should decouple mRoPE coordinate handling from the attention mask generation, preserving the ability to rely on the implicit `is_causal=True` mechanism, which keeps the highly optimized FlashAttention kernel engaged.
2. **Prevent BF16 Precision Collapse in Long-Context RLHF:** The PyTorch `Math` backend is fundamentally unsafe for accumulating Softmax denominators over thousands of tokens (e.g., 8K - 100K) in `bfloat16`. Without the FP32 SRAM accumulators used by fusion kernels, the `Math` backend suffers from severe truncation errors (swamping). Under RLHF losses (like DPO/GRPO/DAPO) which contain exponential amplifiers (`exp(beta * log_probs)`), these errors invariably snowball into catastrophic $10^{28}$ or `NaN` gradients. **Expected behavior:** `transformers` should provide a native `varlen` (variable-length) or `NestedTensors` implementation for `sdpa` that physically truncates padded tokens rather than masking them with `-3.4e38` in a dense tensor, thereby bypassing the mathematically unstable `bfloat16` accumulations.
3. **Explicit Warning or Fallback to FA2:** Until a native SDPA `varlen` solution is implemented, `transformers` should aggressively warn users when `sdpa` is initialized alongside padding masks on models requiring dense mask materialization. Currently, explicitly setting `attn_implementation="flash_attention_2"` is the _only_ mathematically safe approach, as `Qwen2FlashAttention2` uses `cu_seqlens` to physically drop padding and leverages FP32 registers internally, perfectly stabilizing the RLHF gradients. | open | null | false | 2 | [
"bug"
] | [] | 2026-03-22T16:46:05Z | 2026-03-23T10:05:17Z | null | NONE | null | 20260324T120231Z | 2026-03-24T12:02:31Z | ouroborosscr | 75,082,702 | MDQ6VXNlcjc1MDgyNzAy | User | false |
End of preview. Expand in Data Studio
Transformers PR Slop Dataset
Normalized snapshots of issues, pull requests, comments, reviews, and linkage data from huggingface/transformers.
Files:
issues.parquetpull_requests.parquetcomments.parquetissue_comments.parquet(derived view of issue discussion comments)pr_comments.parquet(derived view of pull request discussion comments)reviews.parquetpr_files.parquetpr_diffs.parquetreview_comments.parquetlinks.parquetevents.parquetnew_contributors.parquetnew-contributors-report.jsonnew-contributors-report.md
Use:
- duplicate PR and issue analysis
- triage and ranking experiments
- eval set creation
Notes:
- updated daily
- latest snapshot:
20260324T215331Z - raw data only; no labels or moderation decisions
- PR metadata, file-level patch hunks, and full unified diffs are included
- new contributor reviewer artifacts are included when generated for the snapshot
- full file contents for changed files are not included
- Downloads last month
- 23