[PoC] HF exporters (#41992)
* style
* revert deepseek
* style
* fix style
* fix lfm cache
* revert idefics
* revert idefics
* revert deepseek vl
* better leaf tensors handling
* revert patchtsmixer
* fix in onnx exporter order and executorch backend preparation
* fix get_rope_index models and chunked vision attn models
* fix annotation
* skip video llama 3
* skip ernie vl as well
* introduce dynamo patches and make prepare_for_export the least impactful possible
* fix after merge
* ultimatevlm fix (precompute stuff)
* fix vlm encoders export through optional precomputed inputs
* more vlms passing (t5 gemma and paddle ocr)
* fix qwen omni and paddle ocr export
* fix docs
* fix torch dep
* annotations
* revert change
* missing licence
* attempt fix styling
* fix torch leak
* fix
* style
* fix
* style
* guard
* same name for torch patches
* fix detectron failure
* fix musicflamingo
* fix
* fix
* support and fix gemma4
* update docs
* update docs
* make qwen omni audio encoders exportable
* docs
* get cu seqlens
* leave paddle ocr
* skip paddle ocr
* fix
* fix layoutlmv2 onnx export
* fix idefics3 on cpu
* fix falcon_mamba and other idefics
* style and docs
* dict decomposition
* remove deferred
* executorch fixes
* dynamo doctests passing *-*
* clean title
* update
* normalize qwen omni
* only pure functions
* patch views for exeuctorch
* better vision dynamic tensors pre computation
* audio functions
* fix
* fix
* define vision modeling utils
* fix docs
* style
* fix repo
* update skips
* fixes and skips
* fix repo
* exportable/compilable paddleocr vl
* onnx parity
* more docs
* style
* move
* avoid lens
* timesfm minicpmv46 canine colqwen2
* fix test
* executorch fixes
* executorch fixes
* more executorch fixes
* more executorch patches
* fix
* more fixes
* Update src/transformers/exporters/exporter_dynamo.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* stadarize the patching and fixing api more
* fix get_encoder is some models
* fix style
* fix qwen omni get decoder and export
* fix xlstm export
* per model type inputs preparation registery
* standarize torch and library (onnx, executorch) patching and fx program and graph fixes
* style
* fix gemma4 unified test
* add export for generation pipeline
* decompose for generation helper function
* style
* onnx fixes
* two pat fix
* vectorized swin attention
* vectorized big bird block sparse attn
* avoid dino timeouts
* big bird and pegasus
* fixed unified gemma4
* sam fixes and simplifications
* fix repo
* fix style
* unify model instance patches with the rest of the patching api
* change in api to allow exporting multiple models with diff configs
* accept dicts
* docs review comments
* address comments
* fix clvp by not using datasets and vectorizing loops
* style
* handle moe per backend and better (e.g. onnx translation)
* better export workflow, no dict constraint on inputs, less enforcing impl in prepare_for_export
* modeling fixes not honoring config.use_cache
* style
* style
* easier to mantain inputs preparation
* fix chunk on cuda only
* onnx fixes
* drop cpu gpu sync in mask update
* lower onnx export time for some models by disabling optimize
* fix executorch
* added version drift warnings
* fix more exeutorch
* remove "legacy" name
* model fixes for export compatibility
* add minicpmv vision utils
* can compile
* make ssms fully compilable
* more ssms
* address new comments
* fast ci fixes
* address comments
* fix minimax
* fix the last falcon fast ci failure
* more comments addressed
* attention
* reference attn
* cleanup
* use can_return_tuple in patchmixer
* proper gqa expalanation
* avoid 5D attention
* granite export fix
* address modeling comments
* update
* address more comments, unify linear attention and make more models support full graph compile
* style
* style
* modeling changes for attn impl compatibility
* no need for patch
* fix
* patch is kenrels available
* fixes
* fix lfm2 conv errors
* diffllama is exportable now
* enable more models
* address most of anton's review comments
* address cyril's comments
* revert conv rename
* simpler modular
* keep max_batch_size check
* fixes
* fix moshi cuda graphs
* fix
* added layer type class attr in mixers
* fix repo
* fix
* address comments
* address comments
* address comments
* Merge branch 'main' into hf-exporters-models
* fix merge
* claude review findings
* docs comments
* address comments
* style
* claude review
* small fix caught by copilot
* misc fixes after merge
* onnx fixes
* fix cpu sdpa tripping
* only on cpu
* fix merge
* fix
* address comments
* more comments
* nemotron asr
* fix m3
* style
* add tests for utils
* utils testing
* style
* update deprecation
* trim tests
* style
* skips for executorch
---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Anton Vlasjuk <73884904+vasqu@users.noreply.github.com>
