Transformers 4.53 support, SmolLM3 and Fix old Transformers support#2319
Conversation
| if is_transformers_version(">=", "4.53"): | ||
| from transformers.integrations.executorch import sdpa_mask_without_vmap | ||
| from transformers.masking_utils import AttentionMaskInterface | ||
|
|
||
| AttentionMaskInterface.register("sdpa", sdpa_mask_without_vmap) |
There was a problem hiding this comment.
using the patching spec here doesn't work because the original object is referenced in a dictionary in the AttentionMaskInterface
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
echarlaix
left a comment
echarlaix
left a comment
There was a problem hiding this comment.
Great work @IlyasMoutawwakil thanks a lot !
| def __init_subclass__(cls, **kwargs): | ||
| super().__init_subclass__(**kwargs) | ||
| logger.warning( | ||
| "The `ExportConfig` class is deprecated and will be removed in a future version. " |
There was a problem hiding this comment.
|
Would be nice if gemma3n(multimodal) support is also added along with this PR. |
|
Hi @bil-ash ! unfortunately it won't be that simple because optimum's onnx exporter doesn't support multimodal decoders yet. |
IlyasMoutawwakil
changed the title
|
hi, is there an eta for the merging of this PR? I'd like to use the |
attention vmap patch as in here huggingface#2319
…r normalized config in ort modeling
IlyasMoutawwakil
changed the title
5 participants
What does this PR do?
Fixes # (issue)
Before submitting
Who can review?