Conversation
|
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. |
| TasksManager._CUSTOM_CLASSES[("pt", "idefics3", "image-text-to-text")] = ( | ||
| "transformers", | ||
| "AutoModelForImageTextToText", | ||
| ) | ||
| TasksManager._CUSTOM_CLASSES[("pt", "smolvlm", "image-text-to-text")] = ( | ||
| "transformers", | ||
| "AutoModelForImageTextToText", | ||
| ) |
There was a problem hiding this comment.
redundant ("image-text-to-text" -> AutoModelForImageTextToText) so removing https://github.com/huggingface/optimum/blob/114647f51404f9747410afc287a8079f1f16389b/optimum/exporters/tasks.py#L157
echarlaix
force-pushed
the
transformers-v5
branch
from
01124fa to
7907878
Compare
| "__module.layers.27.mlp.up_proj/aten::linear/MatMul", | ||
| "__module.layers.27.mlp.gate_proj/aten::linear/MatMul", | ||
| ], | ||
| "validate": False, |
There was a problem hiding this comment.
this is needed as for models exported with transformers >= v5 https://huggingface.co/Qwen/Qwen3-Embedding-0.6B/blob/main/config.json#L25 will be loaded in bfloat16 and patched https://github.com/openvinotoolkit/openvino/blob/2026.0.1/src/bindings/python/src/openvino/frontend/pytorch/patch_model.py, for these models we need to use a different default ignored scope config (aten -> ov_ext) depending on what tranformers version was used during the export phase, otherwise quantization is failing at https://github.com/openvinotoolkit/nncf/blob/2f40a6dbc451c6ed025e012dfe65b2d577fde47f/src/nncf/scopes.py#L249. For now one option is to set validate to False as it only concerns Qwen/Qwen3-Embedding-0.6B @rkazants let me know if this works
echarlaix
force-pushed
the
transformers-v5
branch
from
0c5924e to
5033df2
Compare
echarlaix
added
the
openvino-slow
echarlaix
removed
the
openvino-slow
with huggingface/optimum#2408
List of architectures for which are deprecating support for transformers < v5
(remote code modeling incompatible with v5)
(remote code models)
List of architectures for which we removed support for transformers < v5 for which support can be added later on
granitemoeeagle3llavagemma3