Add Top-H decoding (entropy-bounded truncation) as a LogitsWarper for text generation#40837

Merged

gante merged 207 commits intohuggingface:mainfrom

ErfanBaghaei:Top-H-Decoding

Oct 8, 2025

Commits on Sep 9, 2025

Commits on Sep 12, 2025

Commits on Sep 16, 2025

Update logits_process.py removed defaults
ErfanBaghaei
authored

Commits on Sep 22, 2025

Variable H name -> cumulative_entropy
ErfanBaghaei
committed

Commits on Sep 25, 2025

Using torch.distributions.Categorical
ErfanBaghaei
committed
Improve torch_dtype checks (#40808)

authored and
ErfanBaghaei
committed
Add VideoProcessors to auto-backend requirements (#40843)

Cyrilvallez
authored and
ErfanBaghaei
committed
Adds Causal Conv 1D kernel for mamba models (#40765)

MekkCyber
authored and
ErfanBaghaei
committed
Update no split modules in T5Gemma model (#40810)

authored and
ErfanBaghaei
committed
Replace image classification loss functions to `self.loss_function` (#40764)

qubvel
authored and
ErfanBaghaei
committed
Fix the misalignment between the l2norm in GDN of Qwen3-Next and the implementation in the FLA library. (#40842)

authored and
ErfanBaghaei
committed
Fixes for continuous batching (#40828)

remi-or
authored and
ErfanBaghaei
committed
[tests] re-enable aria fast tests (#40846)

gante
authored and
ErfanBaghaei
committed
[SAM2] Fix inconsistent results with original implementation with input boxes (#40800)

yonigozlan
authored and
ErfanBaghaei
committed
[Sam2Video] Fix video inference with batched boxes and add test (#40797)

yonigozlan
authored and
ErfanBaghaei
committed
add: differential privacy research model (#40851)

authored and
ErfanBaghaei
committed
[test] Fix test_eager_matches_sdpa incorrectly skipped (#40852)

eustlb
authored and
ErfanBaghaei
committed
[tests] move generative tests away from `test_modeling_common.py` (#40854)

gante
authored and
ErfanBaghaei
committed
[generate] Always use decoder config to init cache (#40772)

gante
authored and
ErfanBaghaei
committed
Use checkpoint in auto_class_docstring (#40844)

cyyever
authored and
ErfanBaghaei
committed
Fix TrainingArguments.parallelism_config NameError with accelerate<1.10.1 (#40818)

authored and
ErfanBaghaei
committed
Redirect MI355 CI results to dummy dataset (#40862)

ahadnagy
authored and
ErfanBaghaei
committed
[Bug fix #40813] Fix base_model_tp_plan of Starcoder2 model. (#40814)

greg-kwasniewski1
authored and
ErfanBaghaei
committed
[docstrings / type hints] Update outdated annotations for `past_key_values` (#40803)

gante
authored and
ErfanBaghaei
committed
fix florence kwargs (#40826)

SunMarc
authored and
ErfanBaghaei
committed
fix: XIELU act parameters not being casted to correct dtype (#40812)

NanoCode012
authored and
ErfanBaghaei
committed
Update model tags and integration references in bug report (#40881)

ArthurZucker
authored and
ErfanBaghaei
committed
[Qwen3 Next] Use numerically stable `rsqrt` (#40848)

thalahors
authored and
ErfanBaghaei
committed
Adding Support for Qwen3-VL Series (#40795)

authored and
ErfanBaghaei
committed
[`VaultGemma`] Update expectations in integration tests (#40855)

vasqu
authored and
ErfanBaghaei
committed
Fix modular consistency (#40883)

Cyrilvallez
authored and
ErfanBaghaei
committed
🔴 Move variable output controls to `_prepare_generation_config ` (#40715)

manueldeprada
authored and
ErfanBaghaei
committed
Clarify passing is_causal in sdpa_attention_paged_forward (#40838)

cyyever
authored and
ErfanBaghaei
committed
Use torch.expm1 and torch.log1p for better numerical results (#40860)

cyyever
authored and
ErfanBaghaei
committed
Add Fast PromptDepthAnything Processor (#40602)

authored and
ErfanBaghaei
committed
Fix deta loading & dataclass (#40878)

Cyrilvallez
authored and
ErfanBaghaei
committed
Remove dict branch of attention_mask in sdpa_attention_paged_forward (#40882)

cyyever
authored and
ErfanBaghaei
committed
🌐 [i18n-KO] Translated smolvlm.md to Korean (#40414)

authored and
ErfanBaghaei
committed
🌐 [i18n-KO] Translated `imageprocessor.md` to Korean (#39557)

authored and
ErfanBaghaei
committed
[generate] remove docs of a feature that no longer exists (#40895)

gante
authored and
ErfanBaghaei
committed
Make debugging failing tests (check and update expect output values) easier 🔥 (#40727)

ydshieh
authored and
ErfanBaghaei
committed
Fixing the call to kernelize (#40628)

MekkCyber
authored and
ErfanBaghaei
committed
Fix getter regression (#40824)

molbap
authored and
ErfanBaghaei
committed
Fix flaky `Gemma3nAudioFeatureExtractionTest::test_dither` (#40902)

ydshieh
authored and
ErfanBaghaei
committed
[cache] Merge static sliding and static chunked layer (#40893)

Cyrilvallez
authored and
ErfanBaghaei
committed
Harmonize CacheLayer names (#40892)

Cyrilvallez
authored and
ErfanBaghaei
committed
[cache] Only use scalars in `get_mask_sizes` (#40907)

Cyrilvallez
authored and
ErfanBaghaei
committed
Set seed for `Glm4vIntegrationTest` (#40905)

ydshieh
authored and
ErfanBaghaei
committed
Add Olmo3 model (#40778)

2015aroras
authored and
ErfanBaghaei
committed
remove dummy EncodingFast (#40864)

cyyever
authored and
ErfanBaghaei
committed
Improve module name handling for local custom code (#40809)

authored and
ErfanBaghaei
committed
Remove `runner_map` (#40880)

ydshieh
authored and
ErfanBaghaei
committed
disable `test_fast_is_faster_than_slow` (#40909)

ydshieh
authored and
ErfanBaghaei
committed
[gemma3] `Gemma3ForConditionalGeneration` compatible with assisted generation (#40791)

gante
authored and
ErfanBaghaei
committed
[generate] misc fixes (#40906)

gante
authored and
ErfanBaghaei
committed
🔴Make `center_crop` fast equivalent to slow (#40856)

yonigozlan
authored and
ErfanBaghaei
committed
Fix dtype in Paligemma (#40912)

zucchini-nlp
authored and
ErfanBaghaei
committed
[Docs] Adding documentation of MXFP4 Quantization (#40885)

authored and
ErfanBaghaei
committed
Processor load with multi-processing (#40786)

zucchini-nlp
authored and
ErfanBaghaei
committed
[Llama4] Remove `image_sizes` arg and deprecate `vision_feature_layer` (#40832)

yaswanth19
authored and
ErfanBaghaei
committed
Fix #40067: Add dedicated UMT5 support to GGUF loader (config, tokenizer, test) (#40218)

authored and
ErfanBaghaei
committed
[torchao safetensors] renaming get_state_dict function (#40774)

authored and
ErfanBaghaei
committed
Adding activation kernels (#40890)

MekkCyber
authored and
ErfanBaghaei
committed
Minor fix for #40727 (#40929)

ydshieh
authored and
ErfanBaghaei
committed
Add support for Florence-2 training (#40914)

authored and
ErfanBaghaei
committed
Add LongCat-Flash (#40730)

molbap
authored and
ErfanBaghaei
committed
[DOC] Add missing dates in model cards (#40922)

yonigozlan
authored and
ErfanBaghaei
committed
[models] remove unused `import torch.utils.checkpoint` (#40934)

gante
authored and
ErfanBaghaei
committed
Intel CPU dockerfile (#40806)

jiqing-feng
authored and
ErfanBaghaei
committed
docs(i18n): Correct the descriptive text in the README_zh-hans.md (#40941)

lilin-1
authored and
ErfanBaghaei
committed
Fix trainer tests (#40823)

authored and
ErfanBaghaei
committed
Fix `Glm4vMoeIntegrationTest` (#40930)

ydshieh
authored and
ErfanBaghaei
committed
Raise error instead of warning when using meta device in from_pretrained (#40942)

Cyrilvallez
authored and
ErfanBaghaei
committed
Consistent naming for images kwargs (#40834)

zucchini-nlp
authored and
ErfanBaghaei
committed
Remove nested import logic for torchvision (#40940)

yonigozlan
authored and
ErfanBaghaei
committed
Fix `Glm4vModelTest::test_eager_matches_fa2_generate` (#40947)

ydshieh
authored and
ErfanBaghaei
committed
Update expected values for some `test_speculative_generation` (#40949)

ydshieh
authored and
ErfanBaghaei
committed
Standardize audio embedding function name for audio multimodal models (#40919)

jackzhxng
authored and
ErfanBaghaei
committed
Add FlexOlmo model (#40921)

2015aroras
authored and
ErfanBaghaei
committed
Don't list dropout in eager_paged_attention_forward (#40924)

cyyever
authored and
ErfanBaghaei
committed
Update expected values for one more `test_speculative_generation` after #40949 (#40967)

ydshieh
authored and
ErfanBaghaei
committed
FIX(trainer): ensure final checkpoint is saved when resuming training (#40347)

authored and
ErfanBaghaei
committed
Add new model LFM2-VL (#40624)

authored and
ErfanBaghaei
committed
Fix outdated version checks of accelerator (#40969)

cyyever
authored and
ErfanBaghaei
committed
Use `skip_predictor=True` in vjepa2 `get_vision_features` (#40966)

hamishs
authored and
ErfanBaghaei
committed
[Trainer] Fix DP loss (#40799)

authored and
ErfanBaghaei
committed
[timm_wrapper] better handling of "Unknown model" exception in timm (#40951)

harshaljanjani
authored and
ErfanBaghaei
committed
Fix Issue #39030: AutoTokenizer.from_pretrained does not propagate token (#40956)

authored and
ErfanBaghaei
committed
[tests] Really use small models in all fast tests (#40945)

Cyrilvallez
authored and
ErfanBaghaei
committed
Add captured actual outputs to CI artifacts (#40965)

ydshieh
authored and
ErfanBaghaei
committed
Revert change in `compile_friendly_resize` (#40645)

qubvel
authored and
ErfanBaghaei
committed
Track the CI (model) jobs that don't produce test output files (process being killed etc.) (#40981)

ydshieh
authored and
ErfanBaghaei
committed
Using torch.distributions.Categorical
ErfanBaghaei
committed
Remove `set_model_tester_for_less_flaky_tests` (#40982)

Cyrilvallez
authored and
ErfanBaghaei
committed
Benchmarking v2 GH workflows (#40716)

ahadnagy
authored and
ErfanBaghaei
committed
🔴[`Attention`] Bert-based Models Attention Refactor (#38301)

vasqu
authored and
ErfanBaghaei
committed
Remove [[autodoc]] refs to TF/Flax objects (#40996)

Cyrilvallez
authored and
ErfanBaghaei
committed
ENH: Enable readline support for transformers chat (#40911)

BenjaminBossan
authored and
ErfanBaghaei
committed
[testing] test `num_hidden_layers` being small in model tester (#40992)

ydshieh
authored and
ErfanBaghaei
committed
blt wip (#38579)

authored and
ErfanBaghaei
committed
[docs] rm stray tf/flax autodocs references (#40999)

gante
authored and
ErfanBaghaei
committed
[`RMSNorm`] Fix rms norm init for models that center around 1 (#40796)

vasqu
authored and
ErfanBaghaei
committed
Make `EfficientLoFTRModelTest` faster (#41000)

ydshieh
authored and
ErfanBaghaei
committed
Fix typoes in src and tests (#40845)

cyyever
authored and
ErfanBaghaei
committed
Fix more dates in model cards and wrong modalities in _toctree.yml (#40955)

yonigozlan
authored and
ErfanBaghaei
committed
RUFF fix on CI scripts (#40805)

cyyever
authored and
ErfanBaghaei
committed
fix dict like init for ModelOutput (#41002)

SunMarc
authored and
ErfanBaghaei
committed
🚨 [v5] remove generate output retrocompatibility aliases (#40998)

gante
authored and
ErfanBaghaei
committed
[tests] update `test_left_padding_compatibility` (and minimize overwrites) (#40980)

gante
authored and
ErfanBaghaei
committed
Patch more `unittest.case.TestCase.assertXXX` methods (#41008)

ydshieh
authored and
ErfanBaghaei
committed
🚨 [v5] remove deprecated entry point (#40997)

gante
authored and
ErfanBaghaei
committed
🚨 [lightglue] fix: matches order changed because of early stopped indices (#40859)

authored and
ErfanBaghaei
committed
Fix `PhimoeIntegrationTest` (#41007)

ydshieh
authored and
ErfanBaghaei
committed
Fix Glm4v test (#41011)

Cyrilvallez
authored and
ErfanBaghaei
committed
Update after #41007 (#41014)

ydshieh
authored and
ErfanBaghaei
committed
Fix benchmark runner argument name (#41012)

ahadnagy
authored and
ErfanBaghaei
committed
Adding support for Qwen3Omni (#41025)

authored and
ErfanBaghaei
committed
Making compute_loss_func always take priority in Trainer (#40632)

Flakes342
authored and
ErfanBaghaei
committed
Modify Qwen3Omni parameter name since VL changed it (#41045)

authored and
ErfanBaghaei
committed
Fix Qwen video tests (#41049)

zucchini-nlp
authored and
ErfanBaghaei
committed
[testing] Fix `qwen2_audio` (#41018)

ydshieh
authored and
ErfanBaghaei
committed
Fix typing of tuples (#41028)

cyyever
authored and
ErfanBaghaei
committed
Remove optax (#41030)

cyyever
authored and
ErfanBaghaei
committed
Fix typos in English/Chinese documentation (#41031)

cyyever
authored and
ErfanBaghaei
committed
Use torch.autocast (#40975)

cyyever
authored and
ErfanBaghaei
committed
docs: improved RoPE function Docstrings (#41004)

authored and
ErfanBaghaei
committed
Fix condition for emitting warning when generation exceeds max model length (#40775)

yannicks1
authored and
ErfanBaghaei
committed
Fix outdated torch version check (#40925)

cyyever
authored and
ErfanBaghaei
committed
Remove doc of tf and flax (#41029)

cyyever
authored and
ErfanBaghaei
committed
Add Whole Word Masking and Padding Strategy to DataCollatorForLanguageModeling (#39485)

rjgleaton
authored and
ErfanBaghaei
committed
[testing] Fix `seed_oss` (#41052)

authored and
ErfanBaghaei
committed
Remove repeated import (#40937)

cyyever
authored and
ErfanBaghaei
committed
Simplify unnecessary Optional typing (#40839)

cyyever
authored and
ErfanBaghaei
committed
Add write token for uploading benchmark results to the Hub (#41047)

ahadnagy
authored and
ErfanBaghaei
committed
Ci utils (#40978)

remi-or
authored and
ErfanBaghaei
committed
Remove <frameworkcontent> and <pt> tags from documentation (#41055)

authored and
ErfanBaghaei
committed
Fix CI jobs being all red 🔴 (false positive) (#41059)

ydshieh
authored and
ErfanBaghaei
committed
Update quantization CI (#41068)

SunMarc
authored and
ErfanBaghaei
committed
[i18n-bn] Add Bengali language README file (#40935)

saidurpulok
authored and
ErfanBaghaei
committed
Improve documentation and errors in Mamba2-based models (#41063)

mapmeld
authored and
ErfanBaghaei
committed
Update team member list for some CI workflows (#41094)

ydshieh
authored and
ErfanBaghaei
committed
fix crash when using chat to send 2+ request to gptoss (#40536)

sywangyi
authored and
ErfanBaghaei
committed
Minor addition, no split modules for VideoMAEE (#41051)

authored and
ErfanBaghaei
committed
Switch to `python:3.10-slim` for CircleCI docker images (#41067)

ydshieh
authored and
ErfanBaghaei
committed
Fix argument name in benchmarking script (#41086)

ahadnagy
authored and
ErfanBaghaei
committed
Remove mention of TensorFlow/Flax/JAX from English documentation (#41058)

cyyever
authored and
ErfanBaghaei
committed
Fix typos in documentation (#41087)

cyyever
authored and
ErfanBaghaei
committed
Fix typing (#40788)

cyyever
authored and
ErfanBaghaei
committed
Remove unused arguments (#40916)

cyyever
authored and
ErfanBaghaei
committed
Remove tf and flax from Chinese documentation (#41057)

cyyever
authored and
ErfanBaghaei
committed
fix wrong height and width when read video use torchvision (#41091)

Juude
authored and
ErfanBaghaei
committed
docs: Fix Tool Use links and remove dead RAG links (#41104)

RyanMullins
authored and
ErfanBaghaei
committed
🚨 [generate] update paligemma mask updates (and other assisted generation-related fixes) (#40917)

gante
authored and
ErfanBaghaei
committed
[tests] gpt2 + `CausalLMModelTester` (#41003)

gante
authored and
ErfanBaghaei
committed
Fix `_get_test_info` for inherited tests (#41106)

ydshieh
authored and
ErfanBaghaei
committed
Remove bad test skips (#41109)

Cyrilvallez
authored and
ErfanBaghaei
committed
Format empty lines and white space in markdown files. (#41100)

cyyever
authored and
ErfanBaghaei
committed
Update ruff to 0.13.1 + target Python 3.10 + apply fixes (#37809)

authored and
ErfanBaghaei
committed
🚨 [V5] Remove deprecated training arguments (#41017)

cyyever
authored and
ErfanBaghaei
committed
Support loading LFM2 GGUF (#41111)

authored and
ErfanBaghaei
committed
[torchao safetensors] integrate torchao safetensors support with transformers (#40735)

liangel-02
authored and
ErfanBaghaei
committed
[Qwen3-next] Fix dimension mismatch in torch_chunk_gated_delta_rule and torch_recurrent_gated_delta_rule (#40963) (#41036)

authored and
ErfanBaghaei
committed
Fix the error where a keyword argument appearing before *args (#41099)

cyyever
authored and
ErfanBaghaei
committed
Fix broken `` expressions in markdown files (#41113)

cyyever
authored and
ErfanBaghaei
committed
Remove self-assignment (#41062)

authored and
ErfanBaghaei
committed
🚨Refactor: Update text2text generation pipelines to use max_new_tokens… (#40928)

lilin-1
authored and
ErfanBaghaei
committed
Fixed MXFP4 model storage issue (#41118)

YangKai0616
authored and
ErfanBaghaei
committed
Fixed loading LongT5 from legacy checkpoints (#40724)

Szustarol
authored and
ErfanBaghaei
committed
dummy commit (#41133)

ydshieh
authored and
ErfanBaghaei
committed
Fix loading logic flaw with regards to unexpected and missing keys (#40850)

authored and
ErfanBaghaei
committed
Using torch.distributions.Categorical
ErfanBaghaei
committed
Resolving logits_process.py Issues
ErfanBaghaei
committed
style: autoformat with make fixup
ErfanBaghaei
committed
Update logits_process.py removed defaults
ErfanBaghaei
committed
Variable H name -> cumulative_entropy
ErfanBaghaei
committed
Merge branch 'main' into Top-H-Decoding
ErfanBaghaei
authored
Resolving format error
ErfanBaghaei
committed
Correction of the loop variables in logit processor
ErfanBaghaei
committed

Commits on Sep 26, 2025

Vectorized the loop in logits_process
ErfanBaghaei
committed

Commits on Sep 28, 2025

Commits on Sep 30, 2025

Update README.md
souvikku
authored

Commits on Oct 8, 2025