* Rework of the CB example
* Further rework of CB example
* Refactor PA cache, slice on tokens, add debug prints -- WIP
* Slice cache -- WIP
* Added a mechanism to check batched outputs in CB script
* Less logging, debug flag for slice, !better reset! -- WIP
* QOL and safety margins
* Refactor and style
* Better saving of cb example
* Fix
* Fixes and QOL
* Mor einformations about metrics
* Further logging
* Style
* Licenses
* Removed some comments
* Add a slice input flag
* Fix in example
* Added back some open-telemetry deps
* Removed some aux function
* Added FA2 option to example script
* Fixed math (all of it)
* Added a simple example
* Renamed core to classes
* Made allocation of attention mask optionnal
* Style
* Relaxed assumptions on cache_config
* Review compliance
* Style
* Styyyle
* Removed default and added args
* Rebase mishapfix
* Propagate args to TorchExportableModuleForDecoderOnlyLM
* Fix the test I wanted fixed in this PR
* Added some AMD expectation related to cache tests
* draft update two models for now
* batch update all VLMs first
* update some more image processors
* update
* fix a few tests
* just make CI green for now
* fix copies
* update once more
* update
* unskip the test
* fix these two
* fix torchcodec audio loading
* maybe
* yay, i fixed torchcodec installation and now can actually test it
* fix copies deepseek
* make sure the metadata is returrned when users request it
* add docs
* update
* fixup
* Update src/transformers/audio_utils.py
Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com>
* Update src/transformers/models/glm4v/video_processing_glm4v.py
Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com>
* update
* what if we set some metadata attr to `None`
* fix CI
* fix one test
* fix 4 channel test
* fix glm timestemps
* rebase gone wrong
* raise warning once
* fixup
* typo
* fix copies
* ifx smolvlm test
* this is why torch's official benchmark was faster, set threads to `0`
* Apply style fixes
---------
Co-authored-by: Pavel Iakubovskii <qubvel@gmail.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
* initial context_parallel_size support in trainer
* For context parallelism, use AVG instead of SUM to avoid over-accounting tokens
* use parallelism_config.cp_enabled
* add parallelism_config to trainer state
* warn when auto-enabling FSDP
* fix some reviews
* WIP: somewhat matching loss
* Feat: add back nested_gather
* Feat: cleanup
* Fix: raise on non-sdpa attn
* remove context_parallel_size from TrainingArguments
* if we have parallelism_config, we defer to get_state_dict from accelerate
* fix form review
* Feat: add parallelism config support
* Chore: revert some unwanted formatting changes
* Fix: check None
* Check none 2
* Fix: remove duplicate import
* Update src/transformers/trainer.py
Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>
* Update src/transformers/training_args.py
Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>
* Fin
* require accerelate 1.10.1 and higer
---------
Co-authored-by: S1ro1 <matej.sirovatka@gmail.com>
Co-authored-by: Matej Sirovatka <54212263+S1ro1@users.noreply.github.com>
Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>
* Add `tokenizer_kwargs` arg to text generation pipeline.
* chore: re-run CI
* Rename `tokenizer_kwargs` to `tokenizer_encode_kwargs` for text generation pipeline
* Fix `tokenizer_encode_kwargs` doc string.
* Fix note related to `tokenizer _kwargs` in text generation pipeline
---------
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
* add a test
* tempdir
* fix import issue[
* wow I am tired
* properly init
* i am not super familiar with quantizer api :|
* set to TRUE fro now
* full support
* push current changes
* will clean this later but the imports are a shitshow here
* this correctly saves the block and scales but forward seems broken
* quanitze was not correct
* fix storage
* why were bias even included
* finally!
* style
* fix style
* remove print
* lazy import
* up
* not sure what happens this works now?
* holy molly it was not so far
* okay this seems to work!
* workings!!!
* allow save_pretrained to create PR
* Apply suggestions from code review
* fixup
* add deqyabtze fakse as wek
* working new
* fix
* rm swizzle and unswizzle during saving
* rm print
* Update src/transformers/modeling_utils.py
* fix
* style
---------
Co-authored-by: Marc Sun <marc@huggingface.co>
* Fix label smoothing incompatibility with multi-label classification (#40258)
* Improve label smoothing multi-label check based on reviewer feedback
- Move check from LabelSmoother to Trainer.__init__() for better architecture
- Use model.config.problem_type instead of tensor inference for robustness
- Warn and disable smoothing instead of raising error for better UX
- Update test to verify warning behavior
Renamed wer metric variable to wer_metric to avoid naming conflict
with local variable assignment in compute_metrics function.
Co-authored-by: pranam-gf <pranam@goodfin.com>
Fixed 4 instances of the typo "seperator" → "separator" in variable names:
- 2 instances in src/transformers/models/shieldgemma2/convert_shieldgemma2_weights_orbax_to_hf.py
- 2 instances in src/transformers/models/gemma3/convert_gemma3_weights_orbax_to_hf.py
These typos were in variable names used for parsing path components in weight conversion scripts.
🤖 Generated with [Claude Code](https://claude.ai/code)
Co-authored-by: Claude <noreply@anthropic.com>
* fix to the typings which are unmatched to FA function signature
cumulative_seqlens_q/k -> cu_seq_lens_q/k:
- in the FlashAttentionKwargs in modeling_flash_attention_utils
- in the TransformersKwargs in generic
- in the PagedAttentionArgs in continuous_batching
It is **BC**, because they are created in `ContinuousBatchProcessor.setup_static_tensors:L762`, used in `ContinuousBatchingManager._model_forward:L1233` and destroyed with `ContinuousBatchProcessor`
* format changes by ruff
* Update src/transformers/integrations/flash_paged.py
unused function arg in `PagedAttentionCache.update`
Co-authored-by: Anton Vlasjuk <73884904+vasqu@users.noreply.github.com>
* revert continuous_batching signiture, which is more meaningful
---------
Co-authored-by: Anton Vlasjuk <73884904+vasqu@users.noreply.github.com>
* simplify common get/set
* remove some noise
* change some 5 years old modeling utils
* update examples
* fix copies
* revert some changes
* fixes, gah
* format
* move to Mixin
* remove smolvlm specific require grad
* skip
* force defaults
* remodularise some stuff
* remodularise more stuff
* add safety for audio models
* style
* have a correct fallback, you daft donkey
* remove this argh
* change heuristic for audio models
* fixup
* revert
* this works
* this should be explicit
* fix Nth ESM exception
* tryout decoder
* this as well
* revert again
* 🧠
* aaah ESM has two modelings aaah
* broom broom
* format
* wrong copies
* copies
* modular cleanups
* format
* modularities
* wrong mergefix
* seriously
* align with new model
* new model
