mirror of
https://github.com/huggingface/transformers.git
synced 2025-07-14 01:58:22 +06:00
0fe44059ae
* Fork. * RecurrentGemma initial commit. * Updating __init__.py. * Minor modification to how we initialize the cache. Changing how the config specifies the architecture. * Reformat code to 4 spaces. Fixed a few typos. * Fixed the forward pass. Still unclear on the cache? * Fixed the RecurrentGemmaForCausalLM * Minor comment that we might not need attention_mask and output_attention arguments. * Now cache should work as well. * Adding a temporary example to check whether the model generation works. * Adding the tests and updating imports. * Adding the example file missing in the previous commit. * First working example. * Removing .gitignore and reverting parts of __init__. * Re-add .gitignore. * Addressing comments for configuration. * Move mask creation to `_prepare_inputs_for_generation`. * First try at integration tests: 1. AttributeError: 'GriffinCausalLMOutput' object has no attribute 'attentions'. 2. `cache_position` not passed * Transfoering between machines. * Running normal tests. * Minor fix. * More fixes. * Addressing more comments. * Minor fixes. * first stab at cleanup * more refactoring * fix copies and else * renaming and get init to work * fix causal mask creation * update * nit * fix a hell lot of things * updates * update conversion script * make all keys importable * nits * add auto mappings * properly convert ffw_up and down * add scaling * fix generations * for recurrent dtype * update * fix going beyong window * fixup * add missing files * current updates to remove last einops * finish modeling refactor * TADA * fix compile * fix most failing testt ? ? * update tests * refactor and update * update * nits, fixup and update tests * more fixup * nits * fix imports * test format * fixups * nits * tuple typing * fix code quality * add model card * fix doc * skip most generation tests * nits * style * doc fixes * fix pr and check_copies? * last nit * oupsy * Apply suggestions from code review Co-authored-by: Lysandre Debut <hi@lysand.re> * update * Update src/transformers/models/recurrent_gemma/convert_recurrent_gemma_to_hf.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update tests/models/recurrent_gemma/test_modeling_recurrent_gemma.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update tests/models/recurrent_gemma/test_modeling_recurrent_gemma.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update tests/models/recurrent_gemma/test_modeling_recurrent_gemma.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update tests/models/recurrent_gemma/test_modeling_recurrent_gemma.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * update based on review * doc nit * fix quality * quality * fix slow test model path * update default dype * ignore attributes that can be safely ignored in check config attributes * 0lallalala come on * save nit * style * remove to dict update * make sure we can also run in float16 * style --------- Co-authored-by: Pablo Montalvo <39954772+molbap@users.noreply.github.com> Co-authored-by: Aleksandar Botev <botev@google.com> Co-authored-by: Leonard Berrada <lberrada@users.noreply.github.com> Co-authored-by: anushanf <anushanf@google.com> Co-authored-by: botev <botevmg@gmail.com> Co-authored-by: Lysandre Debut <hi@lysand.re> Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
2.0 KiB
2.0 KiB
RecurrentGemma
Overview
The Recurrent Gemma model was proposed in RecurrentGemma: Moving Past Transformers for Efficient Open Language Models by the Griffin, RLHF and Gemma Teams of Google.
The abstract from the paper is the following:
We introduce RecurrentGemma, an open language model which uses Google’s novel Griffin architecture. Griffin combines linear recurrences with local attention to achieve excellent performance on language. It has a fixed-sized state, which reduces memory use and enables efficient inference on long sequences. We provide a pre-trained model with 2B non-embedding parameters, and an instruction tuned variant. Both models achieve comparable performance to Gemma-2B despite being trained on fewer tokens.
Tips:
- The original checkpoints can be converted using the conversion script
src/transformers/models/recurrent_gemma/convert_recurrent_gemma_weights_to_hf.py
This model was contributed by Arthur Zucker.
RecurrentGemmaConfig
autodoc RecurrentGemmaConfig
RecurrentGemmaModel
autodoc RecurrentGemmaModel - forward
RecurrentGemmaForCausalLM
autodoc RecurrentGemmaForCausalLM - forward