Can flagembedding dependancy on transformers upgrade to newer ones?

[firezym]

I have other packages depending on higher versions of transformer such as sentence-transformers, marker-pdf and so on. Can the future releases upgrade transformers version requirements?

flagembedding 1.3.2 has requirement transformers==4.44.2, but you have transformers 4.46.3.
marker-pdf 0.3.10 has requirement transformers<5.0.0,>=4.45.2, but you have transformers 4.44.2.

[Nimtaa]

I’ve also encountered a bug caused by a conflict between transformers==4.44.2 and accelerate==1.2.0. It would be great if you could update the transformers dependency to the latest available version.

[conderls]

File /data/user/usr/anaconda3/envs/py11/lib/python3.11/site-packages/FlagEmbedding/inference/reranker/decoder_only/models/gemma_model.py:56
     54 from .gemma_config import CostWiseGemmaConfig
     55 from transformers.models.gemma2.modeling_gemma2 import Gemma2RMSNorm, Gemma2RotaryEmbedding, rotate_half, apply_rotary_pos_emb
---> 56 from transformers.models.gemma2.modeling_gemma2 import Gemma2MLP, repeat_kv, Gemma2Attention, Gemma2FlashAttention2, Gemma2SdpaAttention, GEMMA2_ATTENTION_CLASSES, Gemma2DecoderLayer, GEMMA2_START_DOCSTRING
     57 from transformers.models.gemma2.modeling_gemma2 import GEMMA2_INPUTS_DOCSTRING
     59 if is_flash_attn_2_available():

ImportError: cannot import name 'GEMMA2_ATTENTION_CLASSES' from 'transformers.models.gemma2.modeling_gemma2'

error raised for transformers==4.47.0

[firezym]

File /data/user/usr/anaconda3/envs/py11/lib/python3.11/site-packages/FlagEmbedding/inference/reranker/decoder_only/models/gemma_model.py:56
     54 from .gemma_config import CostWiseGemmaConfig
     55 from transformers.models.gemma2.modeling_gemma2 import Gemma2RMSNorm, Gemma2RotaryEmbedding, rotate_half, apply_rotary_pos_emb
---> 56 from transformers.models.gemma2.modeling_gemma2 import Gemma2MLP, repeat_kv, Gemma2Attention, Gemma2FlashAttention2, Gemma2SdpaAttention, GEMMA2_ATTENTION_CLASSES, Gemma2DecoderLayer, GEMMA2_START_DOCSTRING
     57 from transformers.models.gemma2.modeling_gemma2 import GEMMA2_INPUTS_DOCSTRING
     59 if is_flash_attn_2_available():

ImportError: cannot import name 'GEMMA2_ATTENTION_CLASSES' from 'transformers.models.gemma2.modeling_gemma2'

error raised for transformers==4.47.0

I have the same problem when I am just importing @545999961

from FlagEmbedding import BGEM3FlagModel

File d:\conda\win\envs\prod\Lib\site-packages\FlagEmbedding\inference\reranker\decoder_only\models\gemma_model.py:56
     [54](file:///D:/conda/win/envs/prod/Lib/site-packages/FlagEmbedding/inference/reranker/decoder_only/models/gemma_model.py:54) from .gemma_config import CostWiseGemmaConfig
     [55](file:///D:/conda/win/envs/prod/Lib/site-packages/FlagEmbedding/inference/reranker/decoder_only/models/gemma_model.py:55) from transformers.models.gemma2.modeling_gemma2 import Gemma2RMSNorm, Gemma2RotaryEmbedding, rotate_half, apply_rotary_pos_emb
---> [56](file:///D:/conda/win/envs/prod/Lib/site-packages/FlagEmbedding/inference/reranker/decoder_only/models/gemma_model.py:56) from transformers.models.gemma2.modeling_gemma2 import Gemma2MLP, repeat_kv, Gemma2Attention, Gemma2FlashAttention2, Gemma2SdpaAttention, GEMMA2_ATTENTION_CLASSES, Gemma2DecoderLayer, GEMMA2_START_DOCSTRING
     [57](file:///D:/conda/win/envs/prod/Lib/site-packages/FlagEmbedding/inference/reranker/decoder_only/models/gemma_model.py:57) from transformers.models.gemma2.modeling_gemma2 import GEMMA2_INPUTS_DOCSTRING
     [59](file:///D:/conda/win/envs/prod/Lib/site-packages/FlagEmbedding/inference/reranker/decoder_only/models/gemma_model.py:59) if is_flash_attn_2_available():

ImportError: cannot import name 'GEMMA2_ATTENTION_CLASSES' from 'transformers.models.gemma2.modeling_gemma2' (d:\conda\win\envs\prod\Lib\site-packages\transformers\models\gemma2\modeling_gemma2.py)

[Nimtaa]

Are there any updated plans or timelines for upgrading to the latest version of transformers? There are security updates in the latest release that we really need to align with: https://github.com/huggingface/transformers/issues/34840

[mxchinegod]

https://github.com/FlagOpen/FlagEmbedding/issues/1356

Adding to the urgency around this. Would be great to have this merged.

[jesnie]

Yes, can we please have some action on this? There even is a pull request for it. You just need to merge it and release it.

[hanhainebula]

Hello, everyone! We've just released version 1.3.4 on PyPi. This update has fixed the dependency issue discussed here. Thanks to @Hypothesis-Z for submitting PR #1343!

[n0isy]

@hanhainebula Thanks, BTW now we don't have any GEMMA2_START_DOCSTRING in the transformers. Which last version of the transformers I can use? Can we fix it (to up to date transformers -OR- add workaround version conditions >= <=)

[545999961]

@n0isy Hello, we have updated the version. Alternatively, you can install transformers==4.46.0.