Elasticsearch is upgraded to PyTorch v2.1.2 for v8.13.0, this change makes Eland use the same version
Due to potential backwards compatibility issues with models traced in PyTorch 2.1.2 running on earlier versions the eland_import_hub_model script will not upload models to a cluster that is earlier than Elasticsearch v8.13.0
ValueError: Attention using SDPA can not be traced with torch.jit.trace when no attention_mask is provided. To solve this issue, please either load your model with the argument `attn_implementation="eager"` or pass an attention_mask input when tracing the model.
The error comes from the decoder layer, in the transformers code the decoder attention mask is not initialised and appears to be added in this PR.
or pass an attention_mask input when tracing the model.
The error message is misleading as the input included attention_mask.
The change was added in transformers v4.36.0 released in December 2023. Rather than work around the problem by injecting the missing parameters the simplest solution is to pin the transformers version to 4.35.2.
Elasticsearch is upgraded to PyTorch v2.1.2 for v8.13.0, this change makes Eland use the same version
Due to potential backwards compatibility issues with models traced in PyTorch 2.1.2 running on earlier versions the
eland_import_hub_model
script will not upload models to a cluster that is earlier than Elasticsearch v8.13.0The tests caught a problem tracing the valhalla/distilbart-mnli-12-6 model:
The error comes from the decoder layer, in the transformers code the decoder attention mask is not initialised and appears to be added in this PR.
The error message is misleading as the input included
attention_mask
.The change was added in transformers v4.36.0 released in December 2023. Rather than work around the problem by injecting the missing parameters the simplest solution is to pin the transformers version to 4.35.2.
Closes #669