[LNL Windows][Test by CD Nightly Wheels] hugging face model - DebertaForQuestionAnswering && DebertaV2ForMaskedLM failed with RuntimeError: value cannot be converted to type at::BFloat16 without overflow

[libohao1201]

🐛 Describe the bug

Huggingface model DebertaForQuestionAnswering && DebertaV2ForMaskedLM failed with RuntimeError: value cannot be converted to type at::BFloat16 without overflow .

Script: python benchmarks\dynamo\huggingface.py --accuracy -d xpu -n10 --inference --backend=eager --cold-start-latency --amp --amp-dtype float16 --only DebertaForQuestionAnswering

Error Log Info:

---

xpu eval DebertaForMaskedLM
Traceback (most recent call last):
File "C:\pt26_ww48\pytorch\benchmarks\dynamo\common.py", line 2732, in validate_model
self.model_iter_fn(model, example_inputs)
File "C:\pt26_ww48\pytorch\benchmarks\dynamo\huggingface.py", line 522, in forward_pass
return mod(**inputs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\transformers\models\deberta\modeling_deberta.py", line 990, in forward
outputs = self.deberta(
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\transformers\models\deberta\modeling_deberta.py", line 798, in forward
encoder_outputs = self.encoder(
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\transformers\models\deberta\modeling_deberta.py", line 609, in forward
hidden_states, att_m = layer_module(
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\transformers\models\deberta\modeling_deberta.py", line 526, in forward
attention_output, att_matrix = self.attention(
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\transformers\models\deberta\modeling_deberta.py", line 461, in forward
self_output, att_matrix = self.self(
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1739, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\torch\nn\modules\module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
File "C:\Users\sdp\pt26_ww48_virtual_env\lib\site-packages\transformers\models\deberta\modeling_deberta.py", line 290, in forward
attention_scores = attention_scores.masked_fill(~(attention_mask), torch.finfo(query_layer.dtype).min)
RuntimeError: value cannot be converted to type at::BFloat16 without overflow

The above exception was the direct cause of the following exception:

Traceback (most recent call last):
File "C:\pt26_ww48\pytorch\benchmarks\dynamo\common.py", line 4856, in run
) = runner.load_model(
File "C:\pt26_ww48\pytorch\benchmarks\dynamo\huggingface.py", line 458, in load_model
self.validate_model(model, example_inputs)
File "C:\pt26_ww48\pytorch\benchmarks\dynamo\common.py", line 2734, in validate_model
raise RuntimeError("Eager run failed") from e
RuntimeError: Eager run failed

eager_fail_to_run

---

Versions

ENV:

Machine	LNL - Win11
stock pytorch	https://github.com/pytorch/pytorch/tree/90b7dcf2c5ee13b892701822f2abbc0e64f5584d pip install --pre torch==2.6.0.dev20241202+xpu torchvision==0.20.0.dev20241202+xpu torchaudio==2.5.0.dev20241202+xpu --index-url https://download.pytorch.org/whl/nightly/xpu
torch-xpu-ops	Commit: bf4bab1
Driver	32.0.101.6314
Conda	python 3.10

[Stonepia]

This is the issue from transformers. See the discussion at huggingface/transformers#35336 .