Cumulative feature and defect updates from recent Transformers PRs by evalstate · Pull Request #42 · evalstate/transformers

evalstate · 2026-04-29T08:56:08Z

Cumulative feature + defect PR mergeability branch

This PR is generated by the all-defects mergeability flow, run with both defect and feature categories enabled. It accumulates recent open huggingface/transformers PRs that could be merged, cherry-picked, patched, or identified as already present on the current base.

Source branch: features-and-defects-750
Base: evalstate/transformers:main
Head: 199afea711
PRs classified: 772
PRs with terminal state: 772
Accepted (merged / applied / already_present): 393
Validation failures reverted: 19
Aborted as impractical/codebase moved on: 213
Skipped categories: 147

Status counts

aborted: 213
already_present: 72
applied: 67
merged: 254
skipped: 147
validation_failed: 19

Category counts

defect: 299
documentation: 97
feature: 326
other: 50

Validation note

The run used the configured lightweight validation profile (compileall, repo checkers, and impacted-test selection when available). Several later batches report validation as unavailable because the baseline/light-validation helper timed out or failed independently of the candidate PR.

The generated report is posted below as a PR comment. Raw local artifacts are:

.mergeability/defect-merge-state.jsonl
.mergeability/pr-classifications.jsonl
all-defects-report.md

# Conflicts: # src/transformers/conversion_mapping.py

# Conflicts: # src/transformers/initialization.py

# Conflicts: # src/transformers/tokenization_utils_tokenizers.py

# Conflicts: # src/transformers/models/colqwen2/modeling_colqwen2.py # src/transformers/models/colqwen2/modular_colqwen2.py # src/transformers/models/paligemma/modeling_paligemma.py

# Conflicts: # tests/test_tokenization_common.py

# Conflicts: # src/transformers/models/pi0/image_processing_pi0.py

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

# Conflicts: # src/transformers/models/auto/configuration_auto.py # src/transformers/models/auto/image_processing_auto.py # src/transformers/models/auto/modeling_auto.py

# Conflicts: # src/transformers/models/colqwen2/modeling_colqwen2.py # src/transformers/models/colqwen2/modular_colqwen2.py

# Conflicts: # src/transformers/integrations/flex_attention.py # src/transformers/utils/generic.py

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

# Conflicts: # src/transformers/models/qwen3_5/modeling_qwen3_5.py

# Conflicts: # src/transformers/conversion_mapping.py # src/transformers/models/auto/configuration_auto.py # src/transformers/models/auto/image_processing_auto.py # utils/check_repo.py

# Conflicts: # src/transformers/pipelines/audio_classification.py # src/transformers/pipelines/automatic_speech_recognition.py # tests/pipelines/test_pipelines_automatic_speech_recognition.py

# Conflicts: # src/transformers/utils/generic.py

Applied from upstream PR huggingface#39103 (1d85c39) with local patch because the direct PR merge conflicts with current Gemma3n v5/strict config docstring structure.

# Conflicts: # src/transformers/loss/loss_utils.py

# Conflicts: # src/transformers/models/kosmos2/modeling_kosmos2.py

# Conflicts: # src/transformers/quantizers/quantizer_bnb_4bit.py # src/transformers/quantizers/quantizer_bnb_8bit.py

# Conflicts: # src/transformers/models/llama4/convert_llama4_weights_to_hf.py

evalstate · 2026-04-29T19:41:10Z

Generated mergeability report

All defects cumulative merge report

Run context

Repo: huggingface/transformers
Worktree: /home/ssmith/source/mergeability-test/.mergeability/defect-worktrees/features-and-defects-750
Branch: features-and-defects-750
Base ref: upstream/main (a8f43eca15b8d1c63deb33f6b97dfab30419e5da when created)
PR list: .mergeability/recent-prs.jsonl (open PRs)
Candidate categories: defect, feature
Validation commands:
- . .mergeability/validation-venv/bin/activate && PYTHONPATH=src python -m compileall -q src/transformers
- . .mergeability/validation-venv/bin/activate && PYTHONPATH=src python utils/checkers.py ruff_check,ruff_format,init_isort,sort_auto_mappings
- . .mergeability/validation-venv/bin/activate && .mergeability/run-light-validation.sh

Record counts

Recent PRs in input list: 750
Classification records: 772
Merge-state records: 772
Recent PRs with terminal state: 750 / 750

Cumulative classifications

defect: 299
feature: 326
documentation: 97
other: 50

Cumulative merge states

merged: 254
applied: 67
already_present: 72
aborted: 213
validation_failed: 19
skipped: 147

Baseline validation for this turn

No baseline validation was run in this final turn because the only remaining unprocessed PR (huggingface#38786) was classified as documentation and skipped. The “before first merge attempt” requirement was therefore not triggered. The previous turn's baseline remains the latest available baseline: compileall passed, while checker/light validation were unavailable due pre-existing cumulative worktree issues and untracked/generated files blocking tests_fetcher.py.

PRs processed this turn

PR Provide clearer instructions on how to specify target language. huggingface/transformers#38786 — documentation — skipped: documentation-only clarification for MADLAD-400 target language tokens; category is not configured for this cumulative branch.

Newly merged/applied PRs this turn

None.

Validation results this turn

No merge was attempted and no validation was required or run this turn.

Recent merged/applied PRs

Fix audio-related config naming for Gemma3n huggingface/transformers#39103: applied (defect, validation unavailable) — direct merge conflicted in Gemma3n files; applied the small naming fix locally; compileall passed but full validation unavailable due baseline issues.
RFC: refactor causal lm loss to handle lm_head in loss function huggingface/transformers#39047: merged (feature, validation unavailable) — merged with local conflict resolution in loss_utils for current typing style; compileall passed but full validation unavailable due baseline issues.
fix kosmos2 tests huggingface/transformers#39037: merged (defect, validation unavailable) — merged with local conflict resolution adapting the is_causal fix to current cache handling; full validation unavailable due baseline issues.
[WIP] Fix DeepseekV3ModelTest::test_torch_compile_for_training huggingface/transformers#39012: already_present (defect, validation unavailable) — intended DeepSeekV3 change was already present.
Add submodels support check function huggingface/transformers#39009: merged (feature, validation unavailable) — merged cleanly; compileall passed but full validation unavailable due baseline issues.
Use deep copies instead of shallow copies for bbox_embed in GroundingDINO decoder (#37333). huggingface/transformers#38999: already_present (defect, validation unavailable) — current GroundingDINO already contains the intended bbox/embed behavior.
Add support to use config dtype in HybridChunkedCache huggingface/transformers#38908: already_present (defect, validation unavailable) — current HybridChunkedCache already contains the intended dtype behavior.
continue to fix distributed_type from TPU to XLA in LM examples (#38652) huggingface/transformers#38888: merged (defect, validation unavailable) — clean merge; validation unavailable because baseline already fails checkers/light validation.
Allow compile with bnb huggingface/transformers#38886: merged (feature, validation unavailable) — merged with local bnb quantizer conflict resolution; PR-local ruff passed, full validation unavailable due baseline issues.
Llama 4 conversion fix for moe models huggingface/transformers#38884: merged (defect, validation unavailable) — merged with local Llama 4 conversion conflict resolution; PR-local ruff passed, full validation unavailable due baseline issues.

Recent aborted PRs

Refactor gemma3n huggingface/transformers#39084: codebase moved on: current Gemma3n is generated from strict modular/config code and already has conflicting audio naming changes.
Fix deprecated max_size parameter handling in DETR image processors huggingface/transformers#39046: codebase moved on: DETR-family image processors have been rewritten around ImageProcessorKwargs/SizeDict helpers, while the PR edits older processor internals and adds backup/ad-hoc verification files.
Remove return_dict kwarg from all the models huggingface/transformers#38991: codebase moved on: PR performs a sweeping return_dict signature/removal refactor across many model files, but current v5 model files have diverged/generated structure.
Check docstring inside modular files as well huggingface/transformers#38988: codebase moved on: PR combines checker changes with widespread modular/model docstring rewrites from an older generated state.
Remove deprecated max_size support from YOLOS image processor huggingface/transformers#38923: codebase moved on: current YOLOS image processor is generated from modular_yolos and inherited DETR v5 image processor kwargs path, while PR edits older generated code and carries a stray submodule entry.
Fix/deprecate max size conditional detr huggingface/transformers#38893: codebase moved on: Conditional DETR image processors have been split/refactored into torch and PIL v5 processors, while PR edits older max_size/from_dict/preprocess code.
Add SamImageProcessorFast with 4x performance improvement huggingface/transformers#38861: codebase moved on: current SAM image processing uses the v5 backend split and generated/dict mappings, while PR adds an older fast image processor and tuple mappings.
Add MobileViT fast image processor huggingface/transformers#38859: codebase moved on: current MobileViT image processing uses the v5 backend split and dynamic imports, while PR adds an older fast processor implementation and tuple mappings.
Add kwargs support in WhisperForConditionalGeneration huggingface/transformers#38810: codebase moved on: Whisper forward signatures now use v5 annotations, Cache, can_return_tuple, and kwargs forwarding; PR is based on older APIs.
Add Dust3R huggingface/transformers#38805: codebase moved on: Dust3R integration conflicts with current auto-generation/modular model structure and v5 image processor backend.

Resume state

Processed this turn: 1
Current consecutive abort count: 4
Next PR to process: none
Termination reason: complete — all PRs in .mergeability/recent-prs.jsonl have terminal state records.

runStatus:
  stopKind: complete
  processedThisTurn: 1
  nextPr: null
  consecutiveAborts: 4

evalstate · 2026-04-29T19:46:09Z

Feature + defect flow status

Processed terminal records: 772

Status counts

Status	Count
aborted	213
already_present	72
applied	67
merged	254
skipped	147
validation_failed	19

Category counts

Category	Count
defect	299
documentation	97
feature	326
other	50

Merged / applied / already-present records (393)

PR	Category	Status	Method	Validation	Original PR summary / goal	Merge note
huggingface#45692	defect	merged	merge	passed	[Fix Phi4 test] Add option to override image_processor_auto_map with local code when trust_remote_code is True	merged cleanly; validation passed with bounded light pytest targets
huggingface#45691	defect	merged	merge	passed	[serve] cb error	merged cleanly; validation passed with bounded light pytest targets
huggingface#45690	feature	merged	merge	passed	[serve] Support for reasoning	merged with adjacent class conflict resolved between serve error and reasoning helpers; validation passed
huggingface#45687	defect	merged	merge	passed	fix: Made histc_input robust for broader hardware	merged cleanly; validation passed with bounded light pytest targets
huggingface#45686	defect	merged	merge	passed	Fix custom-module copies inheriting read-only permissions	merged cleanly; validation passed with bounded light pytest targets
huggingface#45683	defect	merged	merge	passed	Exclude audio modules from conversion process	merged cleanly; ruff fixed whitespace/formatting in changed file and amended merge commit; validation passed
huggingface#45682	defect	merged	merge	passed	FIX Restore LoRA hotswapping functionality	merged cleanly; validation passed with bounded light pytest targets
huggingface#45681	defect	already_present	none	not_run	Restore TokenizersBackend override for DeepSeek V3/R1 tokenizer dispatch	PR conflicted only because current cumulative branch already contains equivalent TokenizersBackend override and DeepSeek regression coverage; no code change ne…
huggingface#45678	defect	merged	merge	passed	Fix shared config mutation issue in flash_attn_from_config	merged cleanly; validation passed with bounded light pytest targets
huggingface#45671	defect	merged	merge	passed	Update latest revision for Phi-4-multimodal test	merged cleanly; validation passed with bounded light pytest targets
huggingface#45670	defect	merged	merge	passed	[nit] glmasr should be in AutoModelForMultimodalLM	merged cleanly; validation passed with bounded light pytest targets
huggingface#45668	feature	merged	merge	passed	[GGUF] Add support for Qwen3.5 MoE (qwen35moe arch)	merged cleanly; validation passed with bounded light pytest targets
huggingface#45662	defect	merged	merge	passed	Fix EP + FSDP2: experts silently overwritten by rank-0 broadcast	merged cleanly; validation passed with bounded light pytest targets
huggingface#45661	feature	merged	merge	passed	[Weight Converter] More fine-grained mappings on classes, scoping for every transforms (including weight converter)	merged cleanly; validation passed with bounded light pytest targets
huggingface#45654	feature	merged	merge	passed	[CB] Refactor any model-related code in a separate class	merged cleanly; validation passed with bounded light pytest targets
huggingface#45653	feature	already_present	none	not_run	[CB] Better overall script and decode bucketting	PR head is already an ancestor of the cumulative branch after merging PR huggingface#45654; no code change needed
huggingface#45651	feature	merged	merge	passed	[Trainer] Optimize LengthGroupedSampler computation with select_columns and tqdm	merged cleanly; validation passed with bounded light pytest targets
huggingface#45649	defect	merged	merge	passed	Fix OOM regression for FSDP2 + cpu_ram_efficient_loading on large models	merged cleanly; validation passed with bounded light pytest targets
huggingface#45645	defect	merged	merge	passed	Fix xdist collisions for captured_info artifacts and preserve CI debug logs	merged cleanly; validation passed with bounded light pytest targets
huggingface#45643	feature	merged	merge	passed	Add DeepSeek V4	merged with conversion_mapping conflict resolved by preserving existing LlavaModel mapping and adding DeepSeek V4 mapping; ruff fixed import order and validati…
huggingface#45642	defect	merged	merge	passed	Fix trust_remote_code local cache collisions for local models (huggingface#45632)	merged with test-file conflict resolved by preserving prior custom_object_save regression test and adding local-cache collision tests; validation passed
huggingface#45640	feature	merged	merge	passed	🚨🚨🚨 [Trainer] Default to FSDP2, simplify API around fsdp + fsdp_config	merged with trainer FSDP setup conflict resolved by preserving EP ignored-module setup while taking FSDP2 defaults; validation passed
huggingface#45639	defect	already_present	none	not_run	Make patched testing debug logs xdist-safe	PR conflicted only with the previously merged huggingface#45645, which already provides xdist-safe captured_info paths, cleanup, and tests; no code change needed
huggingface#45638	feature	merged	merge	passed	Add Multi-Token Prediction (MTP) support for Qwen3.5	merged cleanly; validation passed with bounded light pytest targets
huggingface#45635	defect	merged	merge	passed	qa: speed up dtype regex weight load + reduce dtype tests to 3 random	merged with prefix-handling conflict resolved by preserving interleaved conversion transforms and applying optimized prefix checks; ruff formatted changed file…
huggingface#45634	feature	merged	merge	passed	DeepGEMM BF16, isolation, refactor	merged with moe conflict resolved by preserving DTensor local unwrapping and using a clamped bias index while keeping sentinel detection; validation passed
huggingface#45630	feature	merged	merge	passed	Add new model: Kimi2-6	merged cleanly; ruff formatted Kimi2-6 changed files and validation passed with bounded light pytest targets
huggingface#45627	defect	merged	merge	passed	Processing Utils: honor pre-built sub-processor kwargs in from_pretrained	merged cleanly; validation passed with bounded light pytest targets
huggingface#45626	feature	merged	merge	passed	[Model] Add PP-FormulaNet Model Support	merged cleanly; validation passed with bounded light pytest targets
huggingface#45621	feature	merged	merge	passed	Better Grouped GEMM + EP	merged cleanly; validation passed with bounded light pytest targets
huggingface#45618	feature	merged	merge	passed	Add MTP speculative decoding via MTPCandidateGenerator	merged cleanly; validation passed with bounded light pytest targets
huggingface#45615	defect	merged	merge	passed	fix(qianfan_ocr): add XPU expectations	merged cleanly; validation passed with bounded light pytest targets
huggingface#45614	defect	merged	merge	passed	Add missing requests dependency to transformers[serving]	merged cleanly; ruff formatted setup.py and validation passed with bounded light pytest targets
huggingface#45613	feature	merged	merge	passed	[New Model] Add MiniCPM3 support	merged with auto-mapping conflicts resolved by preserving MiniCPMV4_6 entries and adding MiniCPM3 entries; validation passed
huggingface#45609	feature	merged	merge	passed	make it possible to ser/deser HF MoE models with torchao	merged cleanly; validation passed with bounded light pytest targets
huggingface#45599	feature	merged	merge	passed	qa: more lazy loading	merged cleanly; validation passed with bounded light pytest targets
huggingface#45597	feature	merged	merge	passed	Add Granite 4.1 Vision (granite4_vision)	merged cleanly; validation passed with bounded light pytest targets
huggingface#45596	defect	merged	merge	passed	fix 2 failed test cases for blt model on XPU	merged cleanly; validation passed with bounded light pytest targets
huggingface#45594	defect	merged	merge	passed	fix(utils): Resolve backbone utils test regressions	merged cleanly; validation passed with bounded light pytest targets
huggingface#45591	defect	merged	merge	passed	[nemotron_h] respect _no_reinit flag on dt_bias and out_proj.weight	merged cleanly; validation passed with bounded light pytest targets
huggingface#45586	feature	merged	merge	passed	Add Audio-Visual Flamingo model	merged cleanly; validation passed with bounded light pytest targets
huggingface#45578	defect	merged	merge	passed	Remove attribute_map from GptOssConfig	merged cleanly; validation passed with bounded light pytest targets
huggingface#45570	defect	merged	merge	passed	Fix whisper long-form generation when eos_token_id is a list	merged cleanly; validation passed with bounded light pytest targets
huggingface#45568	defect	merged	merge	passed	Gemma4: fix failed test cases	merged cleanly; validation passed with bounded light pytest targets
huggingface#45552	defect	merged	merge	passed	Remove warnings for modernbert	merged cleanly; validation passed with bounded light pytest targets
huggingface#45549	defect	merged	merge	passed	fix: apply channel averaging correctly in audio feature extractors	merged cleanly; validation passed with bounded light pytest targets
huggingface#45548	defect	merged	merge	passed	Fix EP + DeepSpeed ZeRO-3 loading via accelerate launch	merged cleanly; removed duplicate has_ep definition caused by overlap with existing branch before rerunning validation, which passed with bounded light pytest …
huggingface#45546	feature	merged	merge	passed	feat: Add GGUF loading support for Llama 4 (text)	clean merge; validation passed
huggingface#45541	defect	merged	merge	passed	Fix local_files_only tokenizer fallback when tokenizer files are missing (Issue 45538)	clean merge; validation passed
huggingface#45524	defect	merged	merge	passed	utils: handle flash_attn missing from importlib packages_distributions without crashing	clean merge; validation passed
huggingface#45523	defect	merged	merge	passed	Fix Seq2SeqLM ExecuTorch export: add encoder_attention_mask to decoder and use static encoder shapes	clean merge; validation passed
huggingface#45512	defect	merged	merge	passed	[OutputRecorder] re.search on layer_name	clean merge; validation passed
huggingface#45497	feature	merged	merge	passed	Add V-JEPA 2.1 inference support	clean merge; validation passed
huggingface#45493	feature	merged	merge	passed	Modularize `ProcessorMixin` into smaller components	clean merge; validation passed
huggingface#45490	feature	merged	merge	passed	Add ctsm model	clean merge; validation passed
huggingface#45487	defect	merged	merge	passed	Fix model parallel issue for altclip model and ChineseClip model	clean merge; validation passed
huggingface#45477	feature	merged	merge	passed	Blockwise mask fn as opt arg in all masking functions	clean merge; validation passed
huggingface#45471	feature	merged	merge	passed	Add EXAONE 4.5 implementations	resolved additive IGNORE_NON_TESTED conflict in utils/check_repo.py; validation passed
huggingface#45438	feature	merged	merge	passed	Add Gemma4ForSequenceClassification	resolved Gemma4 additive conflicts with prior masking changes; validation passed
huggingface#45423	defect	merged	merge	unavailable	Fix void segmentation map label reduction	merge clean; compileall and repo style checks passed; light validation unavailable because baseline run-light timed out after 300s
huggingface#45422	defect	merged	merge	unavailable	Drop `content=None` from messages in `apply_chat_template`	merge clean; compileall and repo style checks passed; light validation unavailable because baseline run-light timed out after 300s
huggingface#45413	defect	merged	merge	unavailable	Fix EtaLogitsWarper on fully masked logits	merge clean; compileall and repo style checks passed; light validation unavailable because baseline run-light timed out after 300s
huggingface#45391	feature	merged	merge	unavailable	audio tester class	merge clean after mechanical ruff import/format fixes amended into merge; compileall and repo style checks passed; light validation unavailable because baselin…
huggingface#45389	defect	merged	merge	unavailable	Require input_ids for repetition penalty	merge clean; compileall and repo style checks passed; light validation unavailable because baseline run-light timed out after 300s
huggingface#45379	defect	applied	cherry-pick	passed	fix(config): add deepstack_visual_indexes to Qwen3_5MoeVisionConfig	cherry-pick clean; compileall and repo style checks passed; light validation selected no pytest targets
huggingface#45378	defect	applied	cherry-pick	passed	fix(mistral): guard ReasoningEffort import for older mistral_common versions	cherry-pick clean; compileall and repo style checks passed; light validation selected no pytest targets
huggingface#45351	defect	applied	cherry-pick	passed	fix(testing_utils): guard get_device_capability with torch.cuda.is_available()	cherry-pick clean; compileall and repo style checks passed; light validation ran 203 tests with 107 skipped
huggingface#45350	feature	already_present	none	not_run	WIP: Add support for Granite4VisionForConditionalGeneration	Granite4Vision support is already present in the cumulative branch; direct merge of the older WIP PR conflicts add/add with existing granite4_vision files and …
huggingface#45346	defect	applied	cherry-pick	passed	Fix Double Application of Softmax for Router Logits in MoE models	cherry-picked PR-only MoE router logits fixes; compileall and repo style checks passed; light validation selected no pytest targets
huggingface#45342	defect	applied	cherry-pick	passed	Use `_keys_to_ignore_on_load_unexpected/missing` recursively from children	cherry-pick clean; compileall and repo style checks passed; light validation ran 203 tests with 107 skipped
huggingface#45333	feature	merged	merge	passed	Add heterogeneous config support (per-layer configuration)	merged heterogeneous per-layer config support cleanly; validation passed
huggingface#45300	defect	already_present	none	not_run	Fix Nemotron-H: add mlp layer type support	Nemotron-H mlp layer type support is already present in current configuration/modeling code; direct PR merge only conflicted with renamed layer_types APIs and …
huggingface#45294	feature	already_present	none	not_run	feat: add Gemma4ForSequenceClassification	Gemma4 sequence classification classes, auto mappings, docs, and tests are already present in the current branch; direct merge conflicted with newer Gemma4 tex…
huggingface#45293	defect	applied	cherry-pick	passed	Fix "AttributeError: NewTokenizer has no attribute special_attribute_present" (Remove `REGISTERED_FAST_ALIASES`)	cherry-picked AutoTokenizer registered fast alias cleanup to avoid unrelated merge-parent history; validation passed
huggingface#45273	defect	applied	cherry-pick	passed	fix: liger unnecessarily materializes logits in VRAM during eval, causing OOM	cherry-picked Liger skip_logits eval OOM fix; validation passed
huggingface#45270	feature	applied	cherry-pick	passed	[Trainer] Support multi-loss component logging	cherry-picked Trainer multi-loss component logging to avoid merge-parent tests_fetcher timeout; validation passed
huggingface#45233	defect	applied	patch	passed	feat: make timesfm2_5 onnx export compatible	applied TimesFM ONNX export compatibility diff as a local patch; compile/checkers passed and light validation selected no pytest targets
huggingface#45221	defect	applied	cherry-pick	passed	user friendly error when loading audio from video	cherry-picked user-friendly video-container audio error and reconciled timeout parameter in audio_utils; validation passed
huggingface#45218	feature	applied	patch	passed	Proposal: Agent-first CLI	applied agent-first CLI proposal as a local patch; compile/checkers passed and light validation selected no pytest targets
huggingface#45202	defect	applied	patch	passed	Fix gemma4 has flash-attention incompatbile head-dim=512	applied minimal Gemma4 FlashAttention disable in modular and generated modeling files; compile/checkers passed and light validation selected no pytest targets
huggingface#45193	defect	applied	cherry-pick	passed	Config can apply pyndatic validation without torch-dependence	cherry-picked config pydantic validation fix and style cleanup; validation passed
huggingface#45170	defect	applied	patch	passed	`layrnorm` -> `layernorm`	applied layernorm typo fix across CLIP-like vision models; validation passed
huggingface#45168	feature	applied	cherry-pick	passed	Update min_lr and max_lr default values to better defaults	cherry-picked scheduler default min_lr/max_lr update after direct merge caused light-validation timeout; validation passed
huggingface#45167	feature	applied	cherry-pick	passed	Add anthropic style of function schema	cherry-picked Anthropic-style tool schema support and return-statement refactor; validation passed
huggingface#45157	feature	applied	cherry-pick	passed	[WIP] PrismML Bonsai model support	cherry-picked Prism q1_0_g128 GGUF support and applied formatter; validation passed
huggingface#45147	defect	applied	cherry-pick	passed	Fix broken HQQ support	cherry-picked HQQ integration fixes without PR merge-parent history; validation passed
huggingface#45134	feature	applied	cherry-pick	passed	Optimize Parakeet feature extraction on CUDA	cherry-picked Parakeet CUDA feature-extraction optimization with small conflict resolution; compile/checkers passed and light validation selected no pytest tar…
huggingface#45128	defect	applied	cherry-pick	passed	Fix: handle future annotations in _process_kwargs_parameters	cherry-picked future-annotations auto_docstring fix and amended import/format cleanup; validation passed
huggingface#45105	defect	already_present	none	not_run	Fix @auto_docstring crash with from future import annotations in _process_kwargs_parameters	same auto_docstring future-annotations guard is already present from later PR huggingface#45128 applied on the cumulative branch
huggingface#45086	defect	already_present	none	not_run	fix AttributeError in _patch_mistral_regex	Mistral regex tokenizer AttributeError fix was already present in the cumulative branch; direct merge produced an empty content change and cherry-pick was empty
huggingface#45082	feature	applied	cherry-pick	passed	[VidEoMT] Update conversion script	cherry-picked VidEoMT converter update; compile/checkers passed and light validation selected no pytest targets
huggingface#45075	feature	merged	merge	passed	Add Deepseek-OCR-2 model	merged with mechanical conversion_mapping conflict resolution preserving colqwen2 mapping and adding DeepSeek-OCR-2 mappings; validation passed
huggingface#45040	defect	merged	merge	passed	Llama3 video fix	merged cleanly; ruff formatted VideoLlama3 test file and validation passed
huggingface#44989	feature	merged	merge	passed	🚨 Distributed training API	merged distributed API prototype scripts cleanly; validation passed with bounded light pytest targets
huggingface#44952	defect	merged	merge	unavailable	Fix: Add correct return behaviour when output_hidden_states=True for CLIP and SIGLIP vision models	merge clean; compileall/checkers passed; light validation unavailable because baseline run-light timed out
huggingface#44940	defect	merged	merge	unavailable	Fix tie_weights skipping logic is not tied to model thread scope	merged with mechanical initialization.py conflict resolution preserving meta_device_safe_creation_ops and adopting scoped tie-weight suppression; compileall/ch…
huggingface#44923	defect	merged	merge	unavailable	fix: avoid unconditional model_info call in _patch_mistral_regex	merged with mechanical tokenizer local/offline conflict resolution and amended ruff formatting; compileall/checkers passed; light validation unavailable becaus…
huggingface#44907	defect	merged	merge	unavailable	Remove unnecessary expand_as in get_placeholder_mask across VLMs	merged with mechanical VLM mask conflicts resolved in ColQwen2/PaliGemma and amended missing Callable import; compileall/checkers passed; light validation unav…
huggingface#44893	defect	merged	merge	unavailable	add `StaticLayer.crop()` to match `DynamicLayer` API	merge clean; compileall/checkers passed; light validation unavailable because baseline run-light timed out
huggingface#44891	feature	merged	merge	unavailable	[Trainer] add MoERouterHealthCallback Callback	merge clean; compileall/checkers passed; light validation unavailable because baseline run-light timed out
huggingface#44889	defect	merged	merge	unavailable	[DeepSpeed] Fix evaluate()/predict() before train()	merge clean; compileall/checkers passed; light validation unavailable because baseline run-light timed out
huggingface#45694	defect	merged	merge	unavailable	Fix train_batch_size and eval_batch_size to respect split_batches config	merged cleanly; amended mechanical ruff formatting in added test; compileall/checkers passed; light validation unavailable because baseline and rerun timed out
huggingface#44836	defect	merged	merge	unavailable	Add cu_seqlens support to OlmoHybridGatedDeltaNet for packed sequences	merge clean; compileall/checkers passed; light validation unavailable because baseline run-light timed out
huggingface#44827	defect	merged	merge	unavailable	Fix Mistral4 tests	merge clean; compileall/checkers passed; light validation unavailable because baseline run-light timed out
huggingface#44793	defect	merged	merge	unavailable	fix(janus): Handle None values in image generation mode	merge clean; compileall/checkers passed; light validation unavailable because baseline run-light timed out
huggingface#44781	defect	merged	merge	unavailable	Fix `_set_model_specific_special_tokens` to accept list-format `extra_special_tokens`	merged with modify/delete conflict resolved by preserving current tests/test_tokenization_common.py while applying tokenizer list-handling fix; compileall/chec…
huggingface#44771	defect	merged	merge	unavailable	wtf	merged with mechanical image processor backend rename conflict resolved by applying SigLIP inheritance to current generated PI0 image processor; compileall/che…
huggingface#44731	defect	merged	merge	unavailable	[Tests] Fix slow video tensor creation from list of numpy arrays in SmolVLM	merge clean; compileall/checkers passed; light validation unavailable because baseline and rerun timed out
huggingface#44724	defect	merged	merge	unavailable	Fix some missing / incorrect entries in auto files	merged with mechanical auto-mapping conflicts resolved by keeping current regenerated mappings and adding mlcd_vision_model special module mapping; compileall/…
huggingface#44713	feature	merged	merge	unavailable	[ColQwen2] Refactor output tracing (issue huggingface#43979)	merged with mechanical ColQwen2 output_hidden_states conflict resolution against current kwargs-based forward signature; compileall/checkers passed; light vali…
huggingface#44697	defect	merged	merge	unavailable	fix: torch_float should return float, not int	merged with mechanical conflicts resolved by preserving current flex_attention helper API, applying torch_float float conversion, and adapting Doge flex LSE ca…
huggingface#44680	feature	merged	merge	unavailable	Allow kernel modules to declare their preferred mask function	merge clean; compileall/checkers passed; light validation unavailable because baseline and rerun timed out
huggingface#44676	defect	merged	merge	unavailable	fix(gpt2): Resolve NaN/Inf issue in lm_head on Python 3.13 with tied weights	merge clean; compileall/checkers passed; light validation unavailable because baseline and rerun timed out
huggingface#45695	feature	merged	merge	unavailable	Support for a new Granite-Speech-Plus model	merge clean; compileall/checkers passed; light validation unavailable because baseline and rerun timed out
huggingface#44664	defect	merged	merge	unavailable	🚨 Generic Sequence Classifier works for multimodal models	merged with Qwen3.5 generated-file conflict resolved to current modular layout and amended mechanical ruff formatting; compileall/checkers passed; light valida…
huggingface#44662	feature	merged	merge	unavailable	[model] Add PenguinVL implementation	merged with registry/conversion/check_repo conflicts resolved against current auto_mappings and conversion mapping APIs; compileall/checkers passed; light vali…
huggingface#44660	defect	merged	merge	unavailable	Fix: avoid late CUDA OOM in load_best_model_at_end with PEFT models	merge clean; compileall/checkers passed; light validation unavailable because baseline and rerun timed out
huggingface#44650	defect	merged	merge	unavailable	Fix Seq2SeqTrainer generation path for decoder-only models	merge clean; compileall/checkers passed; light validation unavailable because baseline and rerun timed out
huggingface#44641	defect	merged	merge	unavailable	Conditinally passing and_mask_function arg to create_causal_mask	merge clean; compileall/checkers passed; light validation unavailable because baseline and rerun timed out
huggingface#44635	feature	merged	merge	unavailable	[Gemma] Modular-friendly buffers	merged with Gemma3n RMSNorm buffer conflict resolved by applying buffer fallback to current modular/generated code; compileall/checkers passed; light validatio…
huggingface#45697	defect	already_present	none	unavailable	fix(testing): check torch.cuda.is_available() before get_device_capability	direct merge would replay unrelated BLT conflicts, but the CUDA-headless crash fix is already present in current get_device_properties via a torch.cuda.is_avai…
huggingface#44626	defect	merged	merge	unavailable	don't break legacy behavior when enforced!	merge clean; compileall/checkers passed after rerunning serially; light validation unavailable because baseline and rerun timed out
huggingface#44615	defect	merged	merge	unavailable	Restore is_torch_fx_available for trust_remote_code backwards compatibility	merge clean; compileall/checkers passed; light validation unavailable because baseline timed out
huggingface#44606	defect	merged	merge	unavailable	optionally override tokenizer class with serialized tokenizer	merge clean; compileall/checkers passed; light validation unavailable because baseline timed out
huggingface#44603	defect	merged	merge	unavailable	fixed dockerfile for arm64 systems	merge clean; compileall/checkers passed; light validation unavailable because baseline timed out
huggingface#44594	feature	merged	merge	unavailable	[Pipeline] Add top_k, label filtering, box_format and score sorting to ObjectDetectionPipeline	merge clean; ruff format applied to changed object_detection.py and amended; compileall/checkers passed; light validation unavailable because baseline timed out
huggingface#44587	defect	merged	merge	unavailable	Fix: Handling fused qkv result tensor slicing for tp sharded qkv weights	merge clean; ruff format applied to changed Falcon file and amended; compileall/checkers passed; light validation unavailable because baseline timed out
huggingface#44585	defect	merged	merge	unavailable	Fix missing rms_norm_eps in DeepseekV3 MLA layernorms	merge clean; compileall/checkers passed; light validation unavailable because baseline timed out
huggingface#44569	feature	merged	merge	unavailable	Add SarvamMLA model (sarvamai/sarvam-105b)	merged with current auto-mappings conflict resolved by adding SarvamMLA to generated auto_mappings and preserving existing PPChart2Table config checker excepti…
huggingface#45699	feature	merged	merge	unavailable	Add FP8 kernel acceleration for compressed-tensors quantized models	merge clean; formatter/import-order applied and amended; compileall/checkers passed; light validation unavailable because baseline timed out and tests_fetcher …
huggingface#44543	defect	merged	merge	unavailable	Fix assistant_masks for multimodal inputs in apply_chat_template	merge conflict in adjacent tests resolved by keeping both tests; compileall/checkers passed; light validation unavailable because baseline and rerun timed out
huggingface#44535	defect	already_present	none	unavailable	Fix crash in Qwen2_5_VLProcessor when using batched input with padding=False	PR targets the old Qwen2_5_VLProcessor.call; current branch has moved mm_token_type_ids generation into ProcessorMixin.create_mm_token_type_ids, which alre…
huggingface#44438	feature	merged	merge	unavailable	Add flashoptim	merge clean; compileall and repository style checks passed; light validation unavailable because baseline and post-merge run-light-validation timed out
huggingface#44408	feature	merged	merge	unavailable	Add option to export encoder hidden states for Granite-speech	merge clean; compileall and repository style checks passed; light validation unavailable because baseline and post-merge run-light-validation timed out
huggingface#44385	defect	merged	merge	unavailable	Fix make check-repo	merge clean; compileall and repository style checks passed; light validation unavailable because baseline and post-merge run-light-validation timed out
huggingface#44369	feature	merged	merge	unavailable	Feature/integrations docs fix	merge completed with one straightforward documentation conflict resolved in favor of PR formatting; compileall and repository style checks passed; light valida…
huggingface#44348	feature	merged	merge	unavailable	Enable MetalConfig to load pre-quantized MLX models from HuggingFace Hub	merge completed with one hub.py conflict resolved by preserving current tqdm_class handling and PR MLX shard fallback; compileall and repository style checks p…
huggingface#44270	defect	merged	merge	unavailable	Add correct typing to custom images_kwargs in ProcessorsKwargs	merge completed with straightforward processor typing conflicts resolved by preserving current processor class names/imports and adding PR images_kwargs annota…
huggingface#44259	feature	merged	merge	unavailable	Async data producer	merge clean; compileall and repository style checks passed after narrow ruff/init-isort fixes amended into the merge; light validation unavailable because base…
huggingface#44257	defect	merged	merge	unavailable	use nanmean for aggregating loss	merge clean; compileall and repository style checks passed; light validation unavailable because baseline and post-merge run-light-validation timed out
huggingface#44228	defect	merged	merge	unavailable	[Quantisation] account for nested tensors from quantisers	merge clean; compileall and repository style checks passed; light validation unavailable because baseline and post-merge run-light-validation timed out
huggingface#44215	feature	merged	merge	unavailable	Add sequence classification capability to Granite models	merge completed with a single GraniteMoeHybrid test import conflict resolved to include the new sequence-classification class while preserving current imports;…
huggingface#44189	defect	merged	merge	unavailable	fix: don't move model to device under other dist train backends	merge completed with one Trainer conflict resolved by applying the PR distributed-backend guards to the current fp16/bf16 full-eval device-placement block; com…
huggingface#44184	feature	merged	merge	unavailable	feat: add OpenAI CircuitGPT core architecture and sparse linear layers	merge clean; compileall and repository style checks passed after narrow init-isort fix amended into the merge; light validation unavailable because baseline an…
huggingface#44171	feature	merged	merge	unavailable	Parakeet tdt	merge completed with straightforward conflicts in hub kernel mapping and processor auto mapping resolved by preserving current entries and adding the PR tdt-lo…
huggingface#44159	feature	already_present	none	not_run	Add SDPA and Flash Attention support for OWL-ViT	OWL-ViT attention backend support is already present on the cumulative branch: OwlViTAttention dispatches through ALL_ATTENTION_FUNCTIONS and OwlViTPreTrainedM…
huggingface#44142	feature	merged	merge	unavailable	[voxtral-realtime] get more perfs!	merge completed with Voxtral Realtime generation conflicts resolved by adding precomputed audio_embeds support while preserving the current past_key_values/pos…
huggingface#44070	feature	merged	merge	unavailable	Add GGUF loading support for Qwen3-Next (qwen3_next) architecture	merge completed with GGUF conflicts resolved by preserving current Qwen3.5/MiniMax/Llama4 mappings and adding Qwen3-Next entries; compileall and repository sty…
huggingface#44059	feature	already_present	none	not_run	[GPT2] Refactor output tracing to use capture_outputs/can_return_tuple decorators	GPT-2 output tracing migration is already present on the cumulative branch: GPT2PreTrainedModel has _can_record_outputs, GPT2Model.forward uses capture_outputs…
huggingface#44056	feature	merged	merge	unavailable	[MPNet] Refactor output tracing using capture_outputs decorator	merge completed with a small MPNet conflict resolved in favor of current config.return_dict handling; compileall and repository style checks passed; light vali…
huggingface#44044	feature	merged	merge	unavailable	Refactor DeBERTa output tracing interface	merge completed with DeBERTa-v2 output-tracing conflicts resolved by keeping the decorator-based TransformersKwargs path and dropping stale manual return_dict/…
huggingface#44030	feature	merged	merge	unavailable	refactor output tracing in dpr	merge completed with DPR conflicts resolved by using the decorator-based output tracing path and dropping stale manual output defaults; compileall and reposito…
huggingface#44029	feature	merged	merge	unavailable	refactor output tracing in rwkv	merge completed with RWKV conflicts resolved by using decorator-based output tracing and dropping stale manual return_dict/default handling; compileall and rep…
huggingface#44028	feature	merged	merge	unavailable	refactor output tracing for superpoint	merge completed with SuperPoint conflicts resolved by using decorator-based output tracing and dropping stale manual return_dict/hidden-state defaults; compile…
huggingface#44027	feature	merged	merge	unavailable	refactor output tracing in	merge completed with speech_encoder_decoder output-tracing conflict resolved by keeping decorator-based output handling and dropping stale manual return_dict d…
huggingface#44026	feature	merged	merge	unavailable	refactor output tracing for	merge completed with vision_encoder_decoder output-tracing conflicts resolved by applying decorator-based output handling while keeping current removal of cach…
huggingface#44025	feature	merged	merge	unavailable	refactor output tracing for depth_anything	merge completed by applying depth_anything decorator-based output handling while keeping current prompt_depth_anything generated capture_outputs implementation…
huggingface#44024	feature	merged	merge	unavailable	Focalnet standardized outputs	merge completed with FocalNet output-tracing conflicts resolved by applying the PR capture_outputs/can_return_tuple path while retaining the current backbone h…
huggingface#44019	feature	merged	merge	unavailable	Refactor resnet to use capture_outputs/can_return_tuple output tracing	merge completed with ResNet output-tracing conflicts resolved by applying capture_outputs/can_return_tuple while retaining current backbone hidden-state filter…
huggingface#44017	feature	merged	merge	unavailable	Refactor output tracing in segformers (huggingface#43979)	merge completed with a SegFormer image-classification output-tracing conflict resolved by dropping stale manual return_dict defaulting; compileall and reposito…
huggingface#44013	feature	merged	merge	unavailable	Ouptut tracing: Standardizing MobileNetv2	merge completed with MobileNetV2 output-tracing conflicts resolved by dropping stale manual return_dict/output_hidden_states defaulting; compileall and reposit…
huggingface#44010	feature	merged	merge	unavailable	[SqueezeBert] Migrate to standardized output collection decorators	merge completed with SqueezeBert output-tracing conflicts resolved by applying the decorator-based PR side; compileall and repository style checks passed; ligh…
huggingface#44002	feature	merged	merge	unavailable	refactor output tracing in `upernet`	merge completed with a UPerNet output-flag conflict resolved by keeping the decorator-based kwargs path; compileall and repository style checks passed; light v…
huggingface#44001	feature	merged	merge	unavailable	refactor output tracing in `univnet`	merge completed with UnivNet return_dict conflict resolved by keeping the decorator-based can_return_tuple path; compileall and repository style checks passed;…
huggingface#44000	feature	merged	merge	unavailable	refactor output tracing in `vision_text_dual_encoder`	merge completed cleanly for vision_text_dual_encoder output tracing; compileall and repository style checks passed; light validation unavailable because baseli…
huggingface#43999	feature	merged	merge	unavailable	refactor output tracing in `mobilenet_v1`	merge completed with MobileNetV1 output-tracing conflicts resolved by keeping the decorator-based return path; compileall and repository style checks passed; l…
huggingface#43998	feature	merged	merge	unavailable	refactor output tracing in `timm_backbone`	merge completed with TimmBackbone output-tracing conflicts resolved by keeping the decorator-based kwargs path and removing stale filter_output_hidden_states/i…
huggingface#43997	feature	merged	merge	unavailable	Migrate RegNet to standardized output tracing	merge completed with RegNet output-tracing conflicts resolved by keeping the decorator-based return path; compileall and repository style checks passed; light …
huggingface#43989	defect	merged	merge	unavailable	Fix AutoVideoProcessor class lookup when torchvision is unavailable	merge clean after resolving AutoVideoProcessor lookup conflict for current string-or-None mapping; compileall and style passed; light validation unavailable du…
huggingface#43967	defect	merged	merge	unavailable	Fix AttributeError in run_classification.py when detecting multi-label data	merge clean; compileall and style passed; light validation unavailable due timeout as in baseline
huggingface#43961	defect	already_present	none	not_run	Replace mutable default arguments with None	intended mutable-default fix is already present on the cumulative branch: Idefics uses immutable tuple defaults and debug_utils no longer has the mutable defau…
huggingface#43915	feature	merged	merge	unavailable	add PaddleOCR-VL conversion	merge clean; fixed mechanical ruff issues in the added conversion script and amended merge commit; compileall/style passed; light validation unavailable due ti…
huggingface#43911	defect	merged	merge	unavailable	add Llama to mapping names in tokenization_auto.py	merge clean; compileall and style passed; light validation unavailable due timeout as in baseline
huggingface#43875	defect	merged	merge	unavailable	Improve handling of QuantizedLayer.reset	merge clean; compileall and style passed; light validation unavailable due timeout as in baseline
huggingface#43863	feature	merged	merge	unavailable	[whisper] allow to pass text/audio specific kwargs	merge clean; compileall and style passed; light validation unavailable due timeout as in baseline
huggingface#43842	defect	already_present	none	not_run	fix(cli): Fix TypeAdapter NameError when pydantic is not installed	serve CLI has since been refactored into lazy serving handlers gated by is_serve_available, and no runtime TypeAdapter annotation remains in the current serve.…
huggingface#43838	feature	merged	merge	unavailable	Qwen3 ASR and Forced Aligner	merge completed with small auto-mapping/check_repo conflicts resolved by retaining both current and Qwen3 ASR entries; compileall/style passed; light validatio…
huggingface#43836	defect	already_present	none	not_run	fix: wrapped TypeAdpater in string literals (for now)	current serve CLI no longer contains runtime TypeAdapter annotations after the serving refactor; this duplicate pydantic-optional import fix is already superse…
huggingface#43833	defect	merged	merge	unavailable	fix: ensure dtype consistency in grouped_mm under autocast	merge clean; compileall and style passed; light validation unavailable due timeout as in baseline
huggingface#43826	defect	already_present	none	not_run	fix: error message of pipeline	current PipelineRegistry.check_task has already removed the stale translation_XX_to_YY fallback and now reports only supported tasks, superseding the PR fix
huggingface#43823	feature	merged	merge	unavailable	Add `facebook/MobileLLM-125M`	merge clean; amended mechanical ruff/import/format fixes in added MobileLLM files; compileall and style passed; light validation unavailable due timeout as in …
huggingface#43816	defect	already_present	none	not_run	fix: add id and resume parameters to SwanLab integration	current SwanLab integration already supports SWANLAB_RUN_ID and SWANLAB_RESUME, with an additional default resume=allow when resume_from_checkpoint is used; no…
huggingface#43779	feature	merged	merge	unavailable	SwanLab: Add support for id and resume arguments in SwanLabCallback	merge clean; compileall and style passed; light validation unavailable due timeout as in baseline
huggingface#43775	defect	merged	merge	unavailable	fix(moe): normalize auxiliary loss by top_k for correct load balancing	merge clean; compileall and repository style checks passed; baseline light validation timed out after 300s
huggingface#43747	defect	merged	merge	unavailable	Remove CompressedLinear support for compressed-tensors > 0.13	merge completed with manual conflict resolution preserving current FP8-kernel guard; compileall and repository style checks passed after narrow ruff format; ba…
huggingface#43663	feature	applied	patch	unavailable	Add _get_signature_columns method to allow custom trainers to override column filtering	applied overridable _get_signature_columns hook; compileall and repository style checks passed; baseline light validation timed out after 300s
huggingface#43654	defect	merged	merge	unavailable	fix(tokenizer): Avert special token property overwrites in batch add_tokens calls	merge completed with additive test conflict resolved by keeping both tests; compileall and repository style checks passed; baseline light validation timed out …
huggingface#43651	feature	merged	merge	unavailable	Add _loss_is_scaled_for_ga to allow custom trainers to control gradient accumulation loss scaling	merge clean; compileall and repository style checks passed; baseline light validation timed out after 300s
huggingface#43636	feature	applied	patch	unavailable	Add _metrics dict to Trainer for custom metric logging	applied custom Trainer _metrics collection; compileall and repository style checks passed; baseline light validation timed out after 300s
huggingface#43613	feature	merged	merge	unavailable	Add Promptable Visual Segmentation pipeline	merge completed with generated pipeline registration/docs conflicts resolved to current generated style; compileall and repository style checks passed; baselin…
huggingface#43612	feature	merged	merge	unavailable	Add Promptable Concept Segmentation pipeline	merge completed with adjacent promptable visual/concept segmentation generated registrations kept together; compileall and repository style checks passed; base…
huggingface#43549	defect	merged	merge	unavailable	[kernels] exception handling for fa kernels	merge clean; compileall and repository style checks passed; baseline light validation timed out after 300s
huggingface#43543	defect	merged	merge	unavailable	Fix fp16 underflow in MoE load balancing loss by enforcing fp32 softmax	merge clean; compileall and repository style checks passed; baseline light validation timed out after 300s
huggingface#43542	defect	applied	patch	unavailable	fix: output router capture wrong router logits in qwen moe models	applied router_logits preservation while keeping fp32 softmax routing weights; compileall and repository style checks passed; baseline light validation timed o…
huggingface#43506	feature	merged	merge	unavailable	Add RishAI model with full transformers integration	merge clean; narrow ruff format applied to new RishAI files and amended into merge commit; compileall and repository style checks passed; baseline light valida…
huggingface#43498	defect	merged	merge	unavailable	fix/backward compatibility for tie_weights	merge clean; narrowed deprecated alias to self.tie_weights and applied ruff fix/format in changed file; compileall and repository style checks passed; baseline…
huggingface#43492	feature	merged	merge	unavailable	Perception Encoder follow up PR	merge completed with current strict config style retained and PE audio/video conversion mapping added; compileall and repository style checks passed; baseline …
huggingface#43484	feature	merged	merge	unavailable	Optimize Ernie 4.5 VL timestamp rendering with cached overlays	merge completed with timestamp overlay cache adapted to current Ernie 4.5 VL Moe class naming; compileall and repository style checks passed; baseline light va…
huggingface#43469	feature	merged	merge	unavailable	argparser: Allow optional bool flags without values	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#43466	defect	merged	merge	unavailable	Fix mask loss to ignore padding areas in object detection	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#43451	feature	merged	merge	unavailable	Add Molmo2	merge completed with conversion mapping conflict resolved and auto mappings sorted/amended; compileall and repository style checks passed; light validation una…
huggingface#43395	defect	merged	merge	unavailable	Fix label truncation for per-sample nested structures in Trainer	merge completed with trainer import conflict resolved; compileall and repository style checks passed; baseline light validation unavailable because tests_fetch…
huggingface#43382	feature	merged	merge	unavailable	Allow Path type in transformers.image_utils.load_image function	merge completed with simple image_utils conflict resolved to retain current imports while adding Path input typing; compileall and repository style checks pass…
huggingface#43378	defect	merged	merge	unavailable	feat(models): Make MimiModel encoding padding-aware to ensure batch-to-individual consistency	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#43363	feature	merged	merge	unavailable	[Improvement] Update `DistributedLengthGroupedSampler` to allow customizing length function	merge completed with sampler conflict resolved by retaining current optimized default length computation while adding custom length_func/mega_batch_mult suppor…
huggingface#43291	defect	merged	merge	unavailable	Fix whisper tests	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#43270	defect	merged	merge	unavailable	fix _retrieve_segment timestamps offset bug	merge clean; fixed PR-local trailing whitespace/formatting and amended merge; compileall and repository style checks passed; baseline light validation unavaila…
huggingface#43254	defect	merged	merge	unavailable	Add supported kwargs to fixed_cross_entropy	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#43251	defect	already_present	none	not_run	Fix(43240): pass kwargs to nn.functional.cross_entropy	same fixed_cross_entropy weight and label_smoothing support already merged from PR 43254; PR head is older and would also remove current loss mappings
huggingface#43212	defect	merged	merge	unavailable	Add regression test for offline tokenizer loading (fixes huggingface#43200)	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#43151	defect	merged	merge	unavailable	Make TF32 tests hardware-aware for PyTorch 2.9+	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#43133	defect	merged	merge	unavailable	Fix flaky SAM-HQ integration tests by adding set_seed	merge completed with SAM-HQ conflicts resolved by applying the positional-embedding sharing fix to current modular/generated files; compileall and repository s…
huggingface#43096	defect	applied	cherry-pick	unavailable	Fix save_pretrained for quantized models with custom serialization	applied the defect-fix commit only with the save_original_format guard resolved against current PEFT handling; direct PR-head merge was avoided because the hea…
huggingface#43094	feature	merged	merge	unavailable	Avoid inline .item() sync in decoder start token check	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#43088	feature	merged	merge	unavailable	Skip attention_mask.all() GPU-CPU sync during generation	merge completed with import conflict resolved against current generation utilities; compileall and repository style checks passed; baseline light validation un…
huggingface#43085	feature	merged	merge	unavailable	Add async_stopping_criteria flag to reduce GPU-CPU syncs during generation	merge completed with generation config/utils conflicts resolved and PR-local unused cur_len lint fixed by amendment; compileall and repository style checks pas…
huggingface#43056	feature	merged	merge	unavailable	Perf: enable pin_memory in DataLoader for CLM no_trainer example	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#43044	feature	merged	merge	unavailable	[SAM3] Enable single-scale input support in Mask Decoder	merge completed with test import conflict resolved against current SAM3 test imports and PR-local Union annotation/formatting fixed by amendment; compileall an…
huggingface#43028	defect	applied	patch	unavailable	Fix default interpolation to BICUBIC for ViT, EfficientNet, PVT	ported the remaining ViTImageProcessor default interpolation change to BICUBIC; compileall and repository style checks passed; light validation unavailable aft…
huggingface#43015	defect	merged	merge	unavailable	FIX: TF32 warning (huggingface#43012)	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#42979	defect	merged	merge	unavailable	Fix dtype mismatch in in modeling_llava_next	merge clean; compileall and repository style checks passed; baseline light validation unavailable because tests_fetcher checkout is blocked by pre-existing unt…
huggingface#42942	defect	applied	patch	unavailable	Fix result retrieval starvation and terminate request-scoped iteration on completion	ported starvation fix by deferring mismatched request results until a matching result is found and stopping request iteration on completion/cancellation; compi…
huggingface#42916	defect	already_present	none	unavailable	Fix: Apply clean_up_tokenization_spaces in TokenizersBackend._decode	current TokenizersBackend._decode already honors clean_up_tokenization_spaces with the branch-specific BPE safeguard, which supersedes the PR implementation; n…
huggingface#42881	defect	merged	merge	unavailable	[GGUF] Add attn_logit_softcapping to Gemma2/Gemma3 config mapping	merge clean; compileall and repository checkers passed, but light validation unavailable due pre-existing validation/test_fetcher failures (baseline KeyError; …
huggingface#42865	defect	merged	merge	unavailable	Raise explicit error when FP8 is requested via from_config	merged with localized resolution in modeling_utils.init; compileall and repository checkers passed, but light validation unavailable due pre-existing valid…
huggingface#42793	defect	merged	merge	unavailable	[Quantization] Fixing last issues for a green 2026 CI hopefully	merge clean; compileall and repository checkers passed, but light validation unavailable due pre-existing validation/test_fetcher checkout failures from untrac…
huggingface#42774	defect	already_present	none	not_run	Fix: add None check for extractors in video_processor_class_from_name	current branch already contains the None guard in video_processor_class_from_name (if extractors is not None and class_name in extractors), so PR fix needed no…
huggingface#42765	feature	merged	merge	unavailable	Add distributed training CI	merged with localized CI marker/test-fetcher resolutions and formatter fixes; compileall and repository checkers passed, but light validation unavailable due p…
huggingface#42717	defect	merged	merge	unavailable	image_transforms: fix tensor annotations	merge clean; compileall and repository checkers passed, but light validation unavailable due pre-existing tests_fetcher checkout blocker from untracked generat…
huggingface#42598	defect	merged	merge	unavailable	[pipeline] fix unwanted import failure	merge clean; compileall and repository checkers passed, but light validation unavailable due pre-existing tests_fetcher checkout blocker from untracked generat…
huggingface#42542	defect	already_present	none	not_run	handle get_input_embeddings() on models like gemma3 gracefully	current branch already contains the intended guard for get_input_embeddings returning None or an object without register_forward_hook in enable_input_require_g…
huggingface#42496	feature	already_present	none	not_run	feat: allow CP with trainer	current trainer already contains context-parallel training support including cp_size helpers, shift_labels preparation, padding to cp_size*2, and accelerator.m…
huggingface#42493	defect	merged	merge	unavailable	fix: remove trailing os sep in local pretrained model path	merged with localized test conflict resolution preserving existing dynamic-module tests and adding trailing-separator coverage; compileall and repository check…
huggingface#42446	defect	merged	merge	unavailable	Fix DataParallel dtype access in smolvlm	merge clean; compileall and repository checkers passed, but light validation unavailable due pre-existing tests_fetcher checkout blocker from untracked generat…
huggingface#42432	feature	merged	merge	unavailable	Add VideoToTextPipeline with smart frame sampling and system prompts	merged with localized pipeline registry conflict resolution and formatter-only amend; compileall and repository checkers passed, but light validation unavailab…
huggingface#42424	defect	merged	merge	unavailable	[WIP] attempt to fix ooms in tests	merged with localized Qwen cache_position conflict resolution; compileall and repository checkers passed, but light validation unavailable due pre-existing tes…
huggingface#42403	feature	already_present	none	not_run	Sam3 release on v4.57.3	current cumulative branch already contains SAM3, SAM3 video, tracker, tracker-video, auto mappings, docs, and tests; PR targets v4.57-release and direct merge …
huggingface#42311	defect	applied	patch	unavailable	Fix: Guard against None num_query_tokens in Blip2Processor (to avoid TypeError)	processor guard applied; compileall and style passed; light validation unavailable (tests_fetcher failed, baseline was also unavailable)
huggingface#42310	feature	merged	merge	unavailable	[WIP] Add moondream3 model	merged cleanly; narrow ruff fixes amended; compileall and style passed; light validation unavailable due baseline untracked-file checkout issue
huggingface#42256	feature	merged	merge	unavailable	Integrate Core AutoAWQ Inference Components into Transformers	empty merge commit created; compileall and style passed; light validation unavailable due baseline untracked-file checkout issue
huggingface#42228	defect	merged	merge	unavailable	Support .to(device) or Device Aware Handling for Segmentation Labels in EOMTImageProcessor huggingface#42205	merged cleanly; compileall and style passed; light validation unavailable due baseline untracked-file checkout issue
huggingface#42134	feature	merged	merge	unavailable	Add AutoMergeAdapters utility for merging multiple LoRA adapters with…	merged cleanly; narrow ruff import-order fix amended; compileall and style passed; light validation unavailable due baseline untracked-file checkout issue
huggingface#42133	defect	merged	merge	unavailable	Fix qwen moe Load balancing loss calculation outside training	merged cleanly; compileall and style passed; light validation unavailable due baseline untracked-file checkout issue
huggingface#42131	feature	already_present	none	unavailable	Mistral Tokenizer Converter Script - Initialization with Pattern Argument	PR intent already present: current MistralConverter call already passes the Tekken tokenizer pattern while using the newer PreTrainedTokenizerFast initializati…

evalstate · 2026-04-29T19:46:11Z

Feature + defect flow status table (part 2/4)

Rejected / not included records (379)

PR	Category	Status	Original PR summary / goal	Rejection reason
huggingface#45679	other	skipped	TST Run fast PEFT tests in normal CI	category not configured for this cumulative branch
huggingface#45667	other	skipped	chore(typing): add ty type checking for 3 pipeline files	category not configured for this cumulative branch
huggingface#45666	feature	validation_failed	Extended n-to-1 kernel fusion via	light validation failed in tests_fetcher after merge: invalid relative type-checking imports in src/transformers/integrations/hub_kernels.py produced dependency key src/transformers/integrations/mode…
huggingface#45664	documentation	skipped	Doc translate to Persian(farsi)	category not configured for this cumulative branch
huggingface#45612	documentation	skipped	[docs] update model cards	category not configured for this cumulative branch
huggingface#45608	documentation	skipped	Python code in model docs	category not configured for this cumulative branch
huggingface#45604	feature	aborted	Agent first cli with skill	PR targets upstream/agent-first-cli, not main; direct merge is clean but would also import the absent agent-first-cli base branch, while the PR-only diff depends on missing src/transformers/cli/agent…
huggingface#45569	feature	aborted	Proper nemotron H and 3 and 2	codebase moved on in cumulative branch: PR restructures Nemotron-H into dense/sparse variants and deletes modular_nemotron_h.py, overlapping the already-merged huggingface#45591 _no_reinit initialization fix in…
huggingface#45550	other	skipped	Add runner selection for mi325 GPU type	category not configured for this cumulative branch
huggingface#45543	other	skipped	ci: OTEL support	category not configured for this cumulative branch
huggingface#45534	feature	aborted	🚨 [ALM] Add base model without head	codebase moved on: ALM base-model PR conflicts across auto mappings and multiple generated/model modular files where current branch has newer model mappings/classes (e.g. Granite4Vision/ForConditiona…
huggingface#45476	other	skipped	[Don't merge] Call CI workflow	category not configured for this cumulative branch
huggingface#45465	documentation	skipped	[docs] contributing	category not configured for this cumulative branch
huggingface#45462	other	skipped	chore(sec): added a handful of security checks	category not configured for this cumulative branch
huggingface#45453	feature	aborted	Draft commit	codebase moved on: TP loading refactor conflicts deeply with current core_model_loading weight-mapping/direct-load flow and existing tensor_parallel/finegrained_fp8 updates; resolving would require r…
huggingface#45452	other	skipped	refactor: replace wildcard imports with explicit imports in model init.py files	category not configured for this cumulative branch
huggingface#45426	feature	aborted	Feature/add axk1	codebase moved on: new-model auto registration conflicts with the current auto mapping structure; integrating AXK1 would require adapting generated/config/model mappings to the newer auto-mappings fl…
huggingface#45421	defect	aborted	Improve nested `base_model_prefix` handling in weight conversion and loading	codebase moved on: nested-prefix changes conflict deeply with the current core_model_loading transform pipeline, which now uses ordered WeightTransform handling rather than the PR's separate renaming…
huggingface#45415	other	skipped	Adds type checking to `src/transformers/*py`	category not configured for this cumulative branch
huggingface#45401	feature	aborted	Add support for Voxtral-4B-TTS-2603 to transformers	codebase moved on: new-model auto configuration registration conflicts with the current generated/auto-mappings structure; a safe resolution would require regenerating/adapting Voxtral TTS against th…
huggingface#45396	feature	aborted	Extract dynamic vision/audio tensors into standalone pure functions	codebase moved on: the shared vision/audio tensor extraction conflicts across multiple actively changed modular/generated multimodal models, including newer Qwen3.5 MTP code and GLM/Qwen3 VL changes;…
huggingface#45382	feature	validation_failed	Add AudioGen (AudioCraft) to MusicGen conversion scripts	compileall and style passed after mechanical ruff format amend, but light validation timed out in run-light-validation after repeated 240-290s attempts
huggingface#45363	feature	aborted	n-to-1 kernel fusion via `KernelConfig`	codebase moved on: the kernel fusion series conflicts in utils/kernel_config.py and hub_kernels after newer KernelConfig/hub-kernel changes; the PR branch also merged main, so direct merge would drag…
huggingface#45360	other	skipped	Replace deprecated `huggingface-cli` references with `hf`	category not configured for this cumulative branch
huggingface#45355	feature	aborted	Add universal phone recognition model - PhoneticXeus	codebase moved on: new-model auto configuration registration conflicts with the current auto-mapping structure; integrating PhoneticXeus would require regenerating/adapting auto mappings and docs aga…
huggingface#45332	feature	aborted	Add heterogeneous model support (per-layer config and modeling)	codebase moved on: heterogeneous modeling changes overlap current cache layer registry handling and masking API additions; resolving would require reconciling per-layer heterogeneous layer_idx logic …
huggingface#45321	defect	validation_failed	Remove references to torchao's AffineQuantizedTensor	compile and repository checkers passed, but light validation timed out twice with no pytest result after merging TorchAO AffineQuantizedTensor removal
huggingface#45317	defect	validation_failed	Fix AttributeError in _patch_mistral_regex when fix_mistral_regex=True	narrow cherry-pick avoided unrelated PR-head history and compile/checkers passed, but selected auto-tokenizer validation failed because PixtralProcessor requires missing torchvision
huggingface#45296	feature	aborted	Add GGUF support to Gemma4 (31B & 26B-A4B) text	codebase moved on: Gemma4 GGUF tensor processor/config conversion conflicts with newer GGUF support added for Llama4/Qwen/GPT-OSS in modeling_gguf_pytorch_utils and ggml tests; integrating would requ…
huggingface#45267	documentation	skipped	Add docstring to FFN.forward in DistilBERT	category not configured for this cumulative branch
huggingface#45254	other	skipped	Fix more integration tests for important models	category not configured for this cumulative branch
huggingface#45244	other	skipped	Let's CI go great	category not configured for this cumulative branch
huggingface#45213	other	skipped	DO NOT MERGE - model creation skill	category not configured for this cumulative branch
huggingface#45189	feature	validation_failed	Add doc test CI workflow reusing existing model job infrastructure	merge was clean but light validation timed out after compile/checkers passed
huggingface#45186	feature	aborted	Add new model: Isaac	codebase moved on: new-model auto mappings and conversion mapping conflict with cumulative branch additions; resolving a 96-commit generated/model addition would require regenerating mappings and bro…
huggingface#45181	feature	aborted	Make the cli a top-level package	codebase moved on: cumulative branch already changed the CLI entrypoint for agentic CLI support, conflicting with the PR split into a transformers_cli package
huggingface#45176	feature	aborted	added efficietvitsam model to HF	codebase moved on: new EfficientViT-SAM auto mappings conflict with cumulative branch model registry changes; safe resolution would require regenerated auto mappings and full model review
huggingface#45153	feature	aborted	[`FA`] Native torch integration	codebase moved on: native torch FlashAttention changes conflict with cumulative branch attention-sink/layer_idx handling and newer split_attention_implementation/import-utils availability APIs; resol…
huggingface#45152	documentation	skipped	[docs] model testing	category not configured for this cumulative branch
huggingface#45149	feature	aborted	DO NOT MERGE adding SAML3-LiteText with a skill, first pass	codebase moved on: Sam3 Lite Text files and docs already exist on the cumulative branch and PR also conflicts with generated auto mappings; title is DO NOT MERGE and resolving add/add generated model…
huggingface#45144	feature	aborted	Add Xiaomi MiMo-V2	codebase moved on: new MiMo-V2 model registration conflicts with cumulative branch docs/model registries and auto mappings; resolving an 85-commit generated model addition would require regeneration …
huggingface#45133	feature	validation_failed	Add sarvam model	merge was clean but ruff reported many non-mechanical undefined names/missing imports in new Sarvam modeling/modular files and configuration all; reset merge
huggingface#45115	feature	aborted	Refactor/nemotron h inherit granitemoehybrid	codebase moved on: Nemotron-H refactor conflicts with current generated/model modular code and auto configuration mappings; the PR replaces local block/cache structure with GraniteMoeHybrid inheritan…
huggingface#45114	documentation	skipped	fix: lets fix all doctests	category not configured for this cumulative branch
huggingface#45113	feature	validation_failed	Add GDS support for safetensors loading	merge conflicts were mechanically resolved and compile/checkers passed, but run-light validation timed out after 300s while baseline light validation passed; reset merge
huggingface#45110	feature	aborted	Add SAM 3.1	codebase moved on: SAM 3.1 model addition conflicts with current generated model registries/auto mappings and existing SAM3 conversion script processor changes; safe integration would require regener…
huggingface#45101	feature	aborted	Adding support for Nandi Models	codebase moved on: Nandi model addition conflicts with the current generated auto configuration registry; integrating the new model would require regeneration/adaptation against the cumulative branch…
huggingface#45097	feature	aborted	Add old InternVL2-1B/2B support to the InternVL conversion script huggingface#45092	codebase moved on: InternVL conversion support conflicts across conversion mappings, current auto image/video registries, InternVL processor token handling, and core loading tests; resolving would re…
huggingface#45077	defect	aborted	fix: pin 50 unpinned actions to commit SHA, extract 1 secret to env var	codebase moved on: workflow hardening conflicts across many GitHub Actions files whose current versions already include newer pinned actions/formatting; resolving 13 workflow conflicts would require …
huggingface#45073	feature	validation_failed	Refactor OwlViT to modular Transformers	merge conflict was mechanically resolved and compile/checkers passed, but light validation timed out after 300s while baseline passed; reset merge
huggingface#45067	feature	validation_failed	feat: trainer resume_from_checkpoint support hub downloads (huggingface#43375)	clean merge and compile/checkers passed, but light validation timed out after 300s while baseline passed; reset merge
huggingface#45064	feature	aborted	refactor: shard checkers	codebase moved on: checker sharding conflicts with current Makefile and utils/checkers.py caching/timing/streaming structure; resolving would require CI checker policy review rather than a safe mecha…
huggingface#45060	defect	validation_failed	Fix PIL backend fallback when torchvision is unavailable	clean merge and compile/checkers passed, but light validation timed out after 300s while baseline passed; reset merge
huggingface#45056	defect	validation_failed	[] needs to be only run on doc	merge conflict was mechanically resolved and compile/checkers passed, but light validation timed out after 300s while baseline passed; reset merge
huggingface#45055	defect	aborted	Save model config in Trainer checkpoints for non-PreTrainedModel models	codebase moved on: Trainer save-path patch conflicts in a region that has been substantially reorganized around current gradient/save helpers; safely placing the config save would require reviewing c…
huggingface#45037	documentation	skipped	add missing colon in custom_attention function signature in attention…	category not configured for this cumulative branch
huggingface#45034	defect	aborted	Pass packed boundary metadata to Qwen3.5 linear-attention fast kernels from data collator	codebase moved on: Qwen3.5 linear-attention packed-metadata changes conflict with current gated-delta cache correctness logic in modular/generated model code and overlapping Qwen3.5 tests; resolving …
huggingface#45028	feature	aborted	TP refactor for FSDP + TP integration	codebase moved on: TP/FSDP refactor targets an older fsdp-vs-ddp branch and direct merge drags 65 commits across many model/configuration files plus temporary scripts; conflicts in core_model_loading…
huggingface#45017	defect	aborted	[WIP][Fix] GLM 5 set `apply_rotary_pos_emb` to `is_neox_style=False` && remove `F.relu()`	codebase moved on: WIP GLM5 DSA fix now conflicts with current finegrained FP8 integration and modular/generated GLM MoE DSA implementation; PR also adds new dsa_kernels/FP8 index path beyond the tit…
huggingface#44981	defect	validation_failed	Trainer: set skip_logits for loss-only eval when liger enabled	merge was clean and compile/checkers passed, but light validation timed out after 300s on the Trainer changes while baseline passed; reset merge
huggingface#44979	feature	validation_failed	Module Fusion API	merge was clean and compile/checkers passed, but light validation timed out after 300s while baseline passed; reset merge
huggingface#44974	feature	aborted	Refactor core_model_loading to support FSDP shard-on-read loading	codebase moved on: FSDP shard-on-read core loading refactor is based on the fsdp-vs-ddp workstream and conflicts with current core_model_loading/modeling_utils plus modeling/core-loading tests; resol…
huggingface#44973	defect	validation_failed	Fix max_seqlen type in vision attention for torch.compile + FA2	merge was clean and compile/checkers passed, but light validation timed out after 300s while baseline passed; reset merge
huggingface#44965	other	skipped	try	category not configured for this cumulative branch
huggingface#44958	defect	validation_failed	fixed import error with PILImageResampling	merged cleanly and mechanical ruff fix was amended; compile/checkers then passed, but light validation timed out after 300s while baseline passed; reset merge
huggingface#44956	feature	aborted	Add HyperCLOVAX SEED Think 14B	codebase moved on: HyperCLOVAX model addition conflicts with current docs toctree, models package init, and generated auto configuration/modeling registries; integrating a new generated model safely …
huggingface#44942	feature	aborted	Add inference time layer fusion optimisations via `PreTrainedModel.from_pretrained(fuse_layers=True)`	codebase moved on: PR adds an older fuse_layers boolean and fusion_mapping implementation that conflicts add/add with the cumulative branch newer fusion_config/fusion_mapping API and modeling_utils i…
huggingface#44875	feature	aborted	refactor: improved the cli server module code organization	codebase moved on: CLI serve has already been reorganized around split serving utilities, reasoning mode, and continuous-batching request-id handling; this older single-file refactor conflicts across…
huggingface#44872	documentation	skipped	Fix: Update outdated sampler comment in generation/utils.py	category not configured for this cumulative branch
huggingface#44830	feature	aborted	Add AudioFlamingoNext model	codebase moved on: AudioFlamingoNext adds a generated model alias on top of older MusicFlamingo/AudioFlamingo3 modular files and conflicts with current docs toctree, model package init, auto registri…
huggingface#44815	defect	aborted	Dequant fix	codebase moved on: finegrained FP8 dequantization code has since been refactored and accumulated static expert/Mistral4 support; the PR conflicts in finegrained_fp8.py after earlier Mistral4 changes,…
huggingface#44794	feature	aborted	Refacto GGUF weight conversion	codebase moved on: GGUF conversion refactor now conflicts with current finegrained FP8 integration, GGUF PyTorch loader, and modeling_utils loading flow; the PR head is a broad stacked branch and res…
huggingface#44775	documentation	skipped	[docs] n-d parallelism	category not configured for this cumulative branch
huggingface#44772	documentation	skipped	bitsandbytes: Update links and docs	category not configured for this cumulative branch
huggingface#44729	defect	aborted	Avoid floating point math for ceil operations	codebase moved on: broad int_div_ceil replacement spans configuration, trainer, tensor-parallel, modeling, and many image/video processors; current branch has since refactored fast image processors i…
huggingface#44722	feature	aborted	Refactor gptj output tracing to use standardized decorators	codebase moved on: GPT-J output tracing refactor is based on a version with cache_position plumbing, while the cumulative branch includes later cache_position removal and current output/return_dict h…
huggingface#44682	feature	aborted	transformers serve + llamacpp	codebase moved on: serve has since been split into serving handlers/utilities with reasoning and continuous-batching request-id changes, while this older llama.cpp integration rewrites the monolithic…
huggingface#44659	documentation	skipped	docs: remove outdated use_diff docstring from DistributedConfig.to_js…	category not configured for this cumulative branch
huggingface#44646	documentation	skipped	Fix typo: seperate -> separate	category not configured for this cumulative branch
huggingface#44642	documentation	skipped	Clarify that causal LM labels are shifted internally	category not configured for this cumulative branch
huggingface#44601	feature	aborted	[Distributed] Add PP support natively	codebase moved on: native pipeline-parallel loading patch targets the older core_model_loading flow where every loaded tensor goes through a mapping object; current branch has since added the direct_…
huggingface#44553	feature	aborted	[] Refactor FA CB kwargs	codebase moved on: FA continuous-batching refactor conflicts in flash_paged.py where current branch already has a newer paged decode fast path, lazy kernel import flow, sliding-window handling, and v…
huggingface#44550	documentation	skipped	Improve clarity and grammar in Auto Classes documentation	category not configured for this cumulative branch

evalstate · 2026-04-29T19:46:13Z

Feature + defect flow status table (part 3/4)

| huggingface#44547 | documentation | skipped | Fix position_ids docstring in modeling_flash_attention_utils.py | category not configured for this cumulative branch |
| huggingface#44517 | feature | validation_failed | Add qwen3 tts | merge was clean but ruff reported undefined all export Qwen3TTSTokenizerMultiCodebookCode2WavConfig in qwen3_tts_tokenizer_multi_codebook configuration; reverted top merge |
| huggingface#44495 | feature | aborted | [Gradient Ckpting] Remove unnecessary attribute definitions | codebase moved on: broad gradient-checkpointing cleanup touches hundreds of generated and modular model files; current branch has many intervening model refactors/new models, producing conflicts acro… |
| huggingface#44467 | feature | aborted | Placeholder tokens update | codebase moved on: placeholder-token update spans tokenizer conversion, tokenizer auto mappings, base/tokenizers backends, and model tokenizers; current branch has intervening tokenizer error-handlin… |
| huggingface#44445 | feature | aborted | Adding support for GraniteDoclingHybrid | codebase moved on: GraniteDoclingHybrid adds a new model using older auto-mapping generated files; current branch has since changed configuration_auto to use generated auto_mappings plus non-standard… |
| huggingface#44420 | documentation | skipped | [docs] distributed training | category not configured for this cumulative branch |
| huggingface#44407 | documentation | skipped | docs: add energy efficiency considerations to bitsandbytes quantization guide | category not configured for this cumulative branch |
| huggingface#44394 | feature | aborted | 🚨🚧 FeatureExtractor → AudioProcessor | codebase moved on: PR is a repository-wide FeatureExtractor to AudioProcessor migration touching core preprocessing/image/audio APIs and generated per-model image/audio processors; current cumulative… |
| huggingface#44375 | feature | aborted | Add RF-DETR | codebase moved on: RF-DETR adds a full new model with generated auto mappings and LW-DETR loss changes; current branch has intervening generated auto-mapping/model-registration changes, causing confl… |
| huggingface#44314 | feature | aborted | add HyperClovaX Vision | codebase moved on: HyperClovaX Vision is a large new multimodal model plus Qwen2.5-VL changes and generated auto mappings; current branch has intervening conversion mapping, auto-mapping, video/image… |
| huggingface#44298 | defect | aborted | Auto detect wrong mapping models | codebase moved on: tokenizer backend auto-detection and SentencePiece/Gemma/T5 tokenizer reconstruction have been substantially refactored on the cumulative branch; the PR overlaps current precompile… |
| huggingface#44264 | feature | aborted | [Moe] Enable aux loss automatically when in training + coef is not 0 | codebase moved on: the MoE auxiliary-loss change touches many generated and modular MoE modeling/configuration files; current branch has intervening MoE model refactors and generated-file changes, pr… |
| huggingface#44252 | feature | aborted | Timm unification continued | codebase moved on: timm unification spans backbone utilities, core/model loading, auto mappings, timm_backbone/timm_wrapper configs and modeling, plus many vision model configs/tests; current cumulat… |
| huggingface#44178 | feature | aborted | Add xcodec2 model | codebase moved on: the new XCodec2 model PR was based on the older auto-configuration mapping layout and conflicts with the cumulative branch where auto mappings have been split into auto_mappings/SP… |
| huggingface#44161 | defect | aborted | Refactor LongT5 to use @capture_outputs and @can_return_tuple decorators for unified output handling (Fixes huggingface#43979) | codebase moved on: the LongT5 output-capturing refactor conflicts throughout modeling_longt5.py with the current cumulative branch; it rewrites forward signatures, output plumbing, and tuple/dict ret… |
| huggingface#44154 | defect | aborted | Refactored vits to match standardized output collection interface | codebase moved on: the VITS output-capturing refactor conflicts in modeling_vits.py and uses an older/non-current capture-output API/import shape, so resolving safely would require reworking the deco… |
| huggingface#44129 | defect | aborted | Refactor SpeechT5 output tracing to standardized output capture | codebase moved on: the SpeechT5 output-capturing migration conflicts broadly across modeling_speecht5.py, including encoder/decoder forward signatures, output recorder plumbing, and multiple task hea… |
| huggingface#44123 | defect | aborted | Avoid device sync in training loss accumulation | codebase moved on: Trainer training-loop structure has been substantially refactored on the cumulative branch, with on-device loss accumulation already partially present; the PR conflicts across chec… |
| huggingface#44116 | defect | aborted | [WIP] [Flaubert] Refactor output tracing to decorator-based interface | codebase moved on: the Flaubert decorator migration conflicts throughout modeling_flaubert.py, including imports, attention output contracts, base model forward plumbing, and multiple wrapper heads; … |
| huggingface#44114 | defect | aborted | Migrate wav2vec2, wav2vec2_conformer, and wav2vec2_bert to standardized output collection decorators | codebase moved on: the wav2vec2-family output-capture migration spans many copied/generated and modular audio model files; current cumulative branch has intervening output-capturing, modular-generati… |
| huggingface#44101 | feature | aborted | [XLM] Refactor output tracing to align with capture_outputs standardized architecture | codebase moved on: the XLM output-capture migration conflicts in both XLM and copied Flaubert modeling files; the cumulative branch already contains newer output-capturing/Flaubert refactors, so safe… |
| huggingface#44098 | feature | aborted | [ViLT] Refactor output handling to align with standardized patterns | codebase moved on: ViLT output handling now conflicts in package init exports and modeling_vilt.py with the cumulative branch output-capture changes; resolving would need a current-tree ViLT migratio… |
| huggingface#44086 | feature | validation_failed | [MGP-STR] Refactor output tracing to use capture_outputs/can_return_tuple decorators | merge was clean but compileall failed with SyntaxError: unmatched ")" in src/transformers/models/mgp_str/modeling_mgp_str.py |
| huggingface#44085 | feature | aborted | Refactor RemBERT to use output tracing decorators | codebase moved on: the PR title says RemBERT but the diff edits GPT-J; modeling_gptj.py already has intervening output-tracing changes on the cumulative branch, causing content conflicts that should … |
| huggingface#44083 | feature | aborted | FSDP2 native support in transformers | codebase moved on: native FSDP2 is a broad distributed/modeling feature touching generation, accelerate/FSDP/MoE integrations, modeling utilities, and common tests; the merge conflicts in tests/test_… |
| huggingface#44076 | feature | aborted | Refectored modeling_imagegpt.py to enable hooks to capture_outputs | codebase moved on: ImageGPT modeling has diverged under the output-capture refactor work; the PR conflicts in modeling_imagegpt.py, and applying it safely requires redoing the capture_outputs/can_ret… |
| huggingface#44074 | feature | aborted | [TextNet] Refactor output tracing using capture_outputs decorator | codebase moved on: TextNet output tracing has diverged in the current cumulative branch; modeling_textnet.py conflicts while tests also changed, so safe integration needs a current-tree TextNet captu… |
| huggingface#44073 | feature | aborted | [VisualBert] Refactor output tracing using capture_outputs and can_return_tuple decorators | codebase moved on: VisualBert output-capture decorator migration conflicts in modeling_visual_bert.py with current output handling changes; safely landing it would require reapplying the migration to… |
| huggingface#44072 | feature | aborted | refactor efficientnet output tracing with @capture_outputs and @can_r… | codebase moved on: EfficientNet output tracing changes conflict in modeling_efficientnet.py after intervening capture_outputs/model output refactors; safe resolution requires a fresh migration agains… |
| huggingface#44071 | feature | aborted | [Refactor] Migrate MPT to standardized output tracing decorators | codebase moved on: MPT standardized output tracing migration conflicts in modeling_mpt.py with current cumulative branch output-handling changes; landing it would require a fresh current-tree decorat… |
| huggingface#44068 | feature | aborted | Refactor GPT-Neo to use and decorators | codebase moved on: GPT-Neo output-capture refactor conflicts throughout modeling_gpt_neo.py with current cache_position and output-handling signatures; safely landing it would require reapplying the … |
| huggingface#44066 | feature | aborted | Refactor GPT-J to use standardized output tracing (huggingface#43979) | codebase moved on: GPT-J/CodeGen standardized output tracing conflicts with the current cache_position-aware forward signatures and output plumbing; safe integration would require reapplying the deco… |
| huggingface#44054 | feature | aborted | Flash mla interface | codebase moved on: the experimental Flash MLA/GLM MoE DSA integration conflicts across hub kernel registration, modeling utilities, FP8 integration, GLM MoE DSA config/model/modular files, and tests;… |
| huggingface#44018 | feature | aborted | Refactor GPT-Neo output tracing to use capture_outputs/can_return_tuple | codebase moved on: GPT-Neo output-tracing refactor conflicts throughout modeling_gpt_neo.py with current cache_position-removal and TransformersKwargs/output-capture signatures; safely landing it wou… |
| huggingface#44015 | feature | aborted | Refactor GPT2-based models to standardized output collection interface | codebase moved on: the GPT2/GPTBigCode/DecisionTransformer decorator migration conflicts across current merge_with_config_defaults, OutputRecorder regex layer names, cache_position-removal changes, T… |
| huggingface#44007 | feature | aborted | [ResNet] Refactor output tracing to decorator-based interface | codebase moved on: ResNet output tracing had already been partially migrated on the cumulative branch, while this older PR also changes RegNet/RT-DETR ResNet copied paths and conflicts with current c… |
| huggingface#44004 | feature | aborted | refactor output tracing for codegen | codebase moved on: CodeGen output-tracing refactor conflicts across attention/model/LM forward signatures with current TransformersKwargs and cache_position-removal changes; resolving would require r… |
| huggingface#44003 | feature | aborted | refactor output tracing in mamba | codebase moved on: Mamba/Falcon-Mamba decorator migration conflicts with current cache_position-removal and output handling in imports, backbone forwards, and LM heads; a safe landing would need a fr… |
| huggingface#43996 | feature | aborted | Refactor FNet and CVT output tracing | codebase moved on: FNet/CvT output-tracing refactor conflicts in many model, pretraining, masked-LM, NSP, classification, and QA forward paths with the cumulative branch partially migrated can_return… |
| huggingface#43995 | feature | aborted | Refactoring falcon model to match standardized output collection interface | codebase moved on: Falcon output-tracing refactor conflicts with cumulative branch changes that removed cache_position plumbing and partially migrated output flag handling; resolving safely would req… |
| huggingface#43973 | feature | aborted | Add lfm2.5 audio | codebase moved on: LFM2 audio adds a new model against the older auto-mapping layout, while the cumulative branch now sources configuration mappings from generated auto_mappings plus current generate… |
| huggingface#43924 | feature | aborted | [] More old mask APIs | codebase moved on: broad old attention-mask API migration conflicts across many model files with cumulative output/signature and attention-mask refactors; resolving would require a fresh coordinated … |
| huggingface#43888 | feature | aborted | Support for BharatGen's Param2MoE model architecture | codebase moved on: Param2MoE adds a new model against the older inline configuration_auto mapping, while the cumulative branch uses generated auto_mappings/current auto-generation; safe integration w… |
| huggingface#43785 | defect | aborted | Fix FSDP_CPU_RAM_EFFICIENT_LOADING (huggingface#43749) | codebase moved on: FSDP CPU RAM loading changes conflict in integrations/fsdp.py with current PEFT/FSDP helper additions, and the PR also applies its core_model_loading early-return inside the conver… |
| huggingface#43757 | defect | aborted | Avoid hard failure for gpt-oss GGUF architecture by falling back to g… | codebase moved on: modeling_gguf_pytorch_utils.py now maps gpt-oss to the dedicated gpt_oss architecture and includes GptOssTensorProcessor support, so the PR fallback to gpt-neox conflicts with and … |
| huggingface#43751 | other | skipped | Fix ruff warnings | category not configured for this cumulative branch |
| huggingface#43743 | feature | aborted | Modular playground | codebase moved on: Persimmon modeling has diverged with current cache/output-capturing and kwargs patterns plus an existing modular_persimmon.py, making the playground branch conflicts broad and not … |
| huggingface#43665 | other | skipped | fix | category not configured for this cumulative branch |
| huggingface#43656 | defect | aborted | Fix TypeAdapter NameError in transformers CLI | codebase moved on: cli/serve.py has been substantially rewritten and no longer uses TypeAdapter or the old annotation-heavy serving implementation, so the PR annotations/import fix is obsolete and co… |
| huggingface#43649 | feature | aborted | Check new failures reporting 5 | codebase moved on: CI self-scheduled and notification/check_bad_commit workflows have diverged, including current actor gating and expanded machine/slice matrix conflicting with the PR new-failure re… |
| huggingface#43532 | other | skipped | [do not merge] Show diff | category not configured for this cumulative branch |
| huggingface#43488 | other | skipped | [don't merge] bad format to check repo bot | category not configured for this cumulative branch |
| huggingface#43448 | feature | aborted | Add Molmo | codebase moved on: Molmo adds generated docs/model registries and auto mappings that now conflict with current generated registry structure after Molmo2 and many later model additions; resolving woul… |
| huggingface#43446 | feature | aborted | [typings] Automatically type decorator return types as tuple \| X | codebase moved on: decorator return typing rewrite touches generated modeling/modular files and repo check wiring across the repository; 62 conflicts across core CI/checker files and many model files… |
| huggingface#43424 | other | skipped | Add test to ensure executorch exportability with dynamic shapes | category not configured for this cumulative branch |
| huggingface#43340 | other | skipped | Claude code skills for transformers-api | category not configured for this cumulative branch |
| huggingface#43333 | other | skipped | Fix typo: interupted -> interrupted | category not configured for this cumulative branch |
| huggingface#43310 | other | skipped | Replace regex with re | category not configured for this cumulative branch |
| huggingface#43297 | feature | aborted | [Feat] Reduces redundant tokenization of tags to accelerate Qwen3VL. | codebase moved on: PR implementation rewrites the older Qwen3VLProcessor.call path and the PR head also contains a merge-from-main commit that would drag broad unrelated history; current qwen3_vl… |
| huggingface#43271 | other | skipped | Fix typo: necesary → necessary | category not configured for this cumulative branch |
| huggingface#43267 | documentation | skipped | Add auto_docstring decorator to Sam3ImageProcessorFast | category not configured for this cumulative branch |
| huggingface#43265 | feature | aborted | Adding Omnilingual ASR models | codebase moved on: Omnilingual ASR draft adds a new model against the older explicit auto-mapping layout, while current auto configuration imports generated auto_mappings and requires the modular/new… |
| huggingface#43249 | feature | aborted | [WIP] Processor moves to in | codebase moved on: PR changes processor device handling on an older tree where fast image processor files were tracked, but this cumulative branch has those generated fast image processor files as pr… |
| huggingface#43246 | defect | aborted | GptOss slow tests | codebase moved on: current GptOss slow tests already use Mxfp4Config(dequantize=not quantized), device-qualified expected-output keys, XPU/CPU skips, and expanded distributed/training cases; the PR e… |
| huggingface#43238 | defect | aborted | Fix ObjectDetectionPipeline batch processing bug huggingface#31356 | codebase moved on: PR head includes a checked-in virtualenv and its object_detection.py patch targets an older postprocess signature, duplicating the method and dropping current top_k/labels/box_form… |
| huggingface#43213 | feature | aborted | feat: allow output_hidden_states and output_attensions to record outputs of specific layers | codebase moved on: PR modifies the old check_model_inputs/output-recording hook path in utils.generic, but current generic.py has replaced that flow with merge_with_config_defaults and a deprecated c… |
| huggingface#43192 | feature | aborted | [Trackio] support trackio gpu logging | codebase moved on: TrackioCallback has since gained static Space freezing, bucket_id handling, changed project/space setup, and docs now state report_to defaults to none; the PR GPU-logging/min-versi… |
| huggingface#43149 | defect | aborted | docs(serving): add minimal Python client examples for chat completion… | codebase moved on: PR mixes serving docs with an older monolithic Serve implementation and tests; current serve code has been refactored into cli/serving handlers and docs moved under docs/source/en/… |
| huggingface#43139 | feature | validation_failed | [perf] optimize whisper GPU performance | repository style/lint failed after merge: PR introduces undefined Optional/Union annotations in Whisper feature extractor plus whitespace issues |
| huggingface#43104 | documentation | skipped | docs: clarify tokenizer decoder behavior in v5 (huggingface#43066) | category not configured for this cumulative branch |
| huggingface#43102 | documentation | skipped | Add CPU vs GPU performance comparison example | category not configured for this cumulative branch |
| huggingface#43077 | other | skipped | compileable=>compilable | category not configured for this cumulative branch |
| huggingface#43063 | documentation | skipped | Improve documentation for SegFormer image processor | category not configured for this cumulative branch |
| huggingface#43036 | documentation | skipped | Docs: fix grammar in Pipeline section | category not configured for this cumulative branch |
| huggingface#43020 | feature | aborted | Add mimo v2 flash | codebase moved on: MiMo-V2-Flash new-model PR targets the older explicit auto-mapping files and provides only a modular model skeleton; current auto configuration imports generated auto_mappings and … |
| huggingface#42982 | feature | aborted | Add HumanV: decoder-only causal LM | codebase moved on: HumanV new-model PR targets older auto configuration/modeling mappings and adds broad auto-class entries while current configuration_auto imports generated auto_mappings; safely in… |
| huggingface#42978 | feature | aborted | Add ViT NEPA | codebase moved on: ViT NEPA new-model PR targets older docs/loss registries and explicit auto configuration/modeling mappings; current branch uses generated auto_mappings and has diverged docs/loss r… |
| huggingface#42976 | other | skipped | Upgrade GitHub Actions to latest versions | category not configured for this cumulative branch |
| huggingface#42975 | other | skipped | Upgrade GitHub Actions for Node 24 compatibility | category not configured for this cumulative branch |
| huggingface#42944 | feature | aborted | [Quantization] From config Quantization for FP8 | codebase moved on: FP8 from_config support conflicts in finegrained_fp8 integration and modeling_utils quantization loading paths that have since diverged on the cumulative branch; resolving would re… |
| huggingface#42919 | feature | aborted | [WIP] Video support in vLLM backend | codebase moved on: vLLM video-token support touches many multimodal processors and Llava OneVision modeling; current branch has diverged processor signatures and modular Qwen2.5-VL code, causing broa… |
| huggingface#42908 | defect | aborted | Fix gguf tokenizers | codebase moved on: GGUF tokenizer fix conflicts in generation utilities, auto tokenization mapping, and tokenizer base loading code that have diverged on the cumulative branch; resolving safely would… |
| huggingface#42900 | defect | aborted | Fix: Set clean_up_tokenization_spaces | codebase moved on: PR changes the tokenizer clean_up_tokenization_spaces default to True, but current branch has deliberately changed the default to False and added a BPE-specific escape hatch; apply… |
| huggingface#42887 | feature | aborted | [Quantization] [Compressed Tensors] Support Transforms, Fix Tests | codebase moved on: compressed-tensors transform support conflicts in the quantizer, optional import detection, and compressed-tensors integration tests; current quantization config/import logic has d… |
| huggingface#42876 | documentation | skipped | Document tensor parallelism configuration with Trainer | category not configured for this cumulative branch |
| huggingface#42829 | feature | aborted | [WIP] End-to-end exportable pipelines (object detection) | codebase moved on: exportable object-detection pipeline changes overlap with fast image processor files that were removed/regenerated on the cumulative branch and with divergent image_transforms, RT-… |
| huggingface#42824 | defect | aborted | Fix torch only support for fast Processors | codebase moved on: MLX fast-processor fix depends on image_processing_utils_fast.py and related fast processor tests that are no longer tracked in the current cumulative branch, while feature_extract… |
| huggingface#42816 | defect | aborted | validate tokenizer components | codebase moved on: tokenizer component validation patch conflicts with the now-reworked tokenization_utils_tokenizers import set and convert_to_native_format flow, which now has optimized tokenizer.j… |
| huggingface#42785 | documentation | skipped | Fixing wrong information in Mimi Docs | category not configured for this cumulative branch |
| huggingface#42781 | feature | aborted | Add VibeVoice Realtime | codebase moved on: VibeVoice Realtime depends on a large model-family addition but the cumulative branch already contains overlapping VibeVoice/VibeVoice acoustic tokenizer docs, generated/modeling f… |
| huggingface#42767 | feature | aborted | fix: add mapping of deepseek_v32 model type | codebase moved on: DeepSeek v3.2 model-family addition conflicts with current generated/auto-mapping structure and newer DeepSeek model registrations (deepseek_ocr2/deepseek_v4); configuration_auto.p… |
| huggingface#42744 | other | skipped | [FP8 Devstral 24B] Repro PR | category not configured for this cumulative branch |
| huggingface#42742 | other | skipped | Remove redundant else in activations.py | category not configured for this cumulative branch |
| huggingface#42706 | other | skipped | Nit parakeet | category not configured for this cumulative branch |
| huggingface#42668 | defect | aborted | More robust processor from pretrained | codebase moved on: processor robustness changes overlap with a newer ProcessorMixin subprocessor/modality refactor already on the cumulative branch (_pop_prebuilt_subprocessors, modality aliases, TYP… |
| huggingface#42665 | feature | aborted | Some optimizations for offloading | codebase moved on: offloading optimization conflicts with a substantially refactored core_model_loading path that now uses LoadStateDictInfo, WeightMapping/direct_param_loads, device-mesh helpers, an… |
| huggingface#42655 | feature | aborted | New Feature: Enabling Speculative Decoding with Batch Size > 1 (If draft and target model share tokenizer) | codebase moved on: batched speculative decoding changes conflict with current assisted-generation APIs, candidate generator typing/signatures, assistant_generation_config handling, cache/position-id … |
| huggingface#42631 | defect | aborted | Make GraniteMoeHybridModel compatible with torch.export | codebase moved on: the GraniteMoeHybrid torch.export fix targets a newer/different forward signature using cache_position, check_model_inputs, HybridMambaAttentionDynamicCache imports, and is_torchdy… |
| huggingface#42588 | documentation | skipped | Document the /v1/models endpoint | category not configured for this cumulative branch |
| huggingface#42572 | documentation | skipped | docs: add doctest for SqueezeBERT | category not configured for this cumulative branch |
| huggingface#42527 | documentation | skipped | Added doctests for SwiftFormer model | category not configured for this cumulative branch |
| huggingface#42521 | defect | aborted | Fix FSDP2 defaulting to version 1 in TrainingArguments; add dynamic plugin param passthrough | codebase moved on: FSDP argument handling was substantially refactored on the cumulative branch to normalize legacy fsdp into fsdp_config, default to fsdp_config['version']=2, and relocate FSDP tests… |
| huggingface#42467 | defect | aborted | Fixes StaticCache Crashes | codebase moved on: StaticCache internals on the cumulative branch now use cumulative_length tensors, separate key/value head dimensions, crop support, and a different StaticSlidingWindowLayer update … |
| huggingface#42461 | feature | aborted | Refactor RMSNorm implementations to use torch.nn.functional.rms_norm | codebase moved on: the RMSNorm refactor touches 65 model files and conflicts broadly with current model implementations that have diverged since the PR; although each conflict is conceptually replaci… |
| huggingface#42453 | feature | aborted | Add SDPA and FlashAttention support to T5 | codebase moved on: T5/MT5 modeling files on the cumulative branch diverge substantially from the PR's older attention-interface refactor, producing many conflicts throughout imports, attention classe… |
| huggingface#42437 | other | skipped | One tok typing | category not configured for this cumulative branch |
| huggingface#42430 | other | skipped | 🚨 Clean up image-text-to-text pipeline | category not configured for this cumulative branch |
| huggingface#42415 | other | skipped | initial clean | category not configured for this cumulative branch |
| huggingface#42413 | feature | aborted | Add chatterbox support | codebase moved on: the new Chatterbox/S3Gen/S3Tokenizer model addition targets the older auto-configuration and repo-checker layout; current cumulative branch uses generated auto_mappings plus newer … |
| huggingface#42412 | other | skipped | Replace Optional and Union typing with | in examples | category not configured for this cumulative branch |
| huggingface#42385 | defect | aborted | Fix weight tying logic between tied_weights_keys and tie_word_embeddings | codebase moved on: PR head contains many upstream/main merges and direct merge conflicted in core_model_loading plus dataclass-migrated model config files; intended tied-weight changes would need man… |
| huggingface#42345 | feature | aborted | GPT-OSS Flash Attention and memory-efficient attention via Native PyTorch SDPA | codebase moved on: GPT-OSS attention now uses ALL_ATTENTION_FUNCTIONS.get_interface and kernelized rotary path; PR adds a bespoke sdpa path against the older API and conflicts in generated and modula… |
| huggingface#42292 | documentation | skipped | docs: clarify recommended usage of max_new_tokens in generate() | category not configured for this cumulative branch |
| huggingface#42277 | documentation | skipped | doc(kernels): update kernels integration documentation | category not configured for this cumulative branch |
| huggingface#42244 | defect | aborted | [core] Fix torchao loading | codebase moved on: PR touches cross-cutting torchao/core model-loading code across hundreds of generated modeling files; direct merge produced 0 conflicts including modeling_utils/core_model_loading/… |
| huggingface#42229 | feature | aborted | Add openpangu_moe model | codebase moved on: new model registration conflicts with current generated auto mapping/toctree structure; configuration_auto.py has been refactored from full OrderedDict definitions to generated map… |
| huggingface#42210 | feature | aborted | [WIP] started adding support for evo2 | codebase moved on: Evo2 new-model registration conflicts with current toctree and generated auto mapping/tokenization layout; resolving would require regenerating mappings and adapting the WIP model … |
| huggingface#42166 | feature | aborted | add internvl_flash model | codebase moved on: InternVL Flash new-model registration conflicts with current generated auto configuration mapping and check_repo model ignore rules; resolving requires current model-generation con… |
| huggingface#42130 | defect | aborted | Try refactoring logging to make type checks pass | codebase moved on: logging.py has since gained lazy tqdm/huggingface_hub imports and a TransformersLogger typing protocol while this PR rewrites logger typing via setLoggerClass; resolving would requ… |
| huggingface#42127 | defect | aborted | Standardize conv len function for audio models | codebase moved on: audio convolution-length standardization now conflicts in multiple generated and modular audio model files; resolving would require propagating/regenerating model copies rather tha… |
| huggingface#42124 | documentation | skipped | 📚 docs(qwen3): add comprehensive usage examples and model details | category not configured for this cumulative branch |
| huggingface#42112 | feature | aborted | Add max_thinking_tokens for reasoning models (issue huggingface#42111) | codebase moved on: max_thinking_tokens generation feature conflicts with current GenerationConfig and generation tests; integrating would require adapting the new logits processor/config semantics to… |
| huggingface#42039 | other | skipped | [WIP] 🚨 clean xcodec 🧼 | category not configured for this cumulative branch |
| huggingface#42000 | documentation | skipped | Fix Mixtral: Docstring uses consistent 'top_k_index' and 'top_k_weights' in MixtralSparseMoeBlock | category not configured for this cumulative branch |
| huggingface#41992 | feature | aborted | [PoC] HF exporters | codebase moved on: broad exporter framework touches many modeling, modular, utility, and documentation files and now conflicts in multiple generated/modular multimodal model files; resolving would re… |
| huggingface#41980 | defect | aborted | Correct type hint in config models | codebase moved on: config rope/type-hint annotations have diverged across many generated and modular model configuration files; resolving would require broad modular regeneration and current typing r… |
| huggingface#41977 | feature | aborted | Add Phi3.5 Vision Model | codebase moved on: new Phi3.5 Vision model registration conflicts with current generated auto configuration/image-processing/modeling mappings; integrating would require regenerating and adapting the… |
| huggingface#41967 | feature | aborted | feat: RoPE-related typing improvements | codebase moved on: RoPE utility typing has changed in current modeling_rope_utils, conflicting with the PR typing refactor; resolution would need reviewing current RoPE parameter schemas rather than … |
| huggingface#41899 | feature | aborted | Testing checkpoint limit changes from PR huggingface#37196 | codebase moved on: checkpoint save-limit feature conflicts with current Trainer, TrainingArguments, and trainer tests; resolving would require reworking new save limit semantics against the current t… |
| huggingface#41886 | feature | aborted | ADD FG-CLIP2 | codebase moved on: FG-CLIP2 adds a full model and collides with current auto-mapping registries; resolving would require regenerating/validating model integration against the current generated auto f… |
| huggingface#41882 | feature | aborted | Support fdma for models with attention bias | codebase moved on: FDMA integration conflicts in attention docs, modeling utilities, testing utilities, and utility exports; resolution would need adapting to current attention/backend registration A… |
| huggingface#41880 | documentation | skipped | Indonesian Language Support for ReadMe | category not configured for this cumulative branch |
| huggingface#41823 | feature | aborted | Lfm2-VL vllm | codebase moved on: PR head is an old v4.57 release branch with ~1489 changed files and direct merge conflicts across CI, docs, generation, auto mappings, image processors, and the current LFM2-VL mod… |
| huggingface#41807 | documentation | skipped | git commit -m "Fix: corrected outdated documentation link in README.md" | category not configured for this cumulative branch |
| huggingface#41800 | documentation | skipped | Increasing clarity | category not configured for this cumulative branch |
| huggingface#41794 | other | skipped | Enable flake8-pie rules | category not configured for this cumulative branch |
| huggingface#41754 | feature | aborted | Add pytree registration for static cache | codebase moved on: StaticCache pytree registration PR was written against older cache layer names and older ExecuTorch export helpers; current code has StaticSlidingWindowLayer/current DynamicCache r… |
| huggingface#41733 | documentation | skipped | transformers CLI documentation issue | category not configured for this cumulative branch |
| huggingface#41721 | defect | aborted | Fix Qwen3-VL Processor flattening multi-image batches (fix huggingface#41709) | codebase moved on: PR edits the old monolithic Qwen3VLProcessor.call, but current branch has moved Qwen3-VL processing to ProcessorMixin helpers such as replace_image_token/replace_video_token wi… |
| huggingface#41710 | documentation | skipped | 🌐 [i18n-KO] Translated main_classes/backbones.md to Korean | category not configured for this cumulative branch |
| huggingface#41693 | feature | aborted | 🚨 Refactor ViT to updated standards | codebase moved on: PR is a broad vision-model standards refactor touching dozens of generated/model files; current cumulative branch already has newer conversion_mapping/test_modeling_common record-o… |
| huggingface#41654 | defect | aborted | Improve LLaMA tokenizer error when vocab is missing: suggest installi… | codebase moved on: PR modifies the old full slow LlamaTokenizer.get_spm_processor implementation, but current v5 branch has reduced tokenization_llama.py to the TokenizersBackend shim with no local S… |
| huggingface#41611 | documentation | skipped | Docs add custom loss example | category not configured for this cumulative branch |
| huggingface#41609 | defect | aborted | Fix gemma gguf tokenizer | codebase moved on: direct merge pulls stacked docs/processor changes and conflicts in tokenization_auto.py and processing_utils.py; the actual Gemma GGUF commit targets the old AutoTokenizer flow and… |
| huggingface#41597 | documentation | skipped | Standardize RoBERTa model card following issue huggingface#36979 | category not configured for this cumulative branch |
| huggingface#41584 | defect | aborted | Add clear error message for missing SentencePiece model in get_spm_processor (fix huggingface#41553) | codebase moved on: PR changes the old slow LlamaTokenizer.get_spm_processor SentencePiece implementation, but current v5 tokenization_llama.py is a TokenizersBackend/BPE shim with no get_spm_processo… |
| huggingface#41565 | documentation | skipped | 🌐 [i18n-KO] Updated perf_train_gpu_many.md | category not configured for this cumulative branch |
| huggingface#41557 | other | skipped | Update modeling_llama4.py | category not configured for this cumulative branch |
| huggingface#41531 | documentation | skipped | 🌐 [i18n-KO] Translated video_processor.md to Korean | category not configured for this cumulative branch |
| huggingface#41527 | documentation | skipped | 🌐 [i18n-KO] Translated selecting.md to Korean | category not configured for this cumulative branch |
| huggingface#41491 | feature | aborted | Add skip_unnecessary_grad_clip to TrainingArguments for optimized gradient clipping | codebase moved on: PR head includes a large merge from upstream and direct merge conflicts in the refactored Trainer/TrainingArguments training loop; the actual gradient-clipping change targets the o… |
| huggingface#45679 | other | skipped | TST Run fast PEFT tests in normal CI | category not configured for this cumulative branch |
| huggingface#41488 | other | skipped | Create 1 | category not configured for this cumulative branch |
| huggingface#41485 | defect | aborted | Fix smolvlm2 dtype mismatch final | codebase moved on: SmolVLM modular/modeling dtype handling already diverged with a newer self.dtype/StopIteration fallback, conflicting with the PR's inputs_embeds.dtype change in generated and modul… |
| huggingface#41419 | feature | aborted | First QAT for Finegrained FP8 | codebase moved on: QAT FP8 changes conflict in central lazy imports, import_utils optional dependency helpers, and finegrained_fp8 integration now changed substantially; safe resolution would require… |
| huggingface#41406 | other | skipped | 🚨 [v5] Remove deprecated cache classes | category not configured for this cumulative branch |
| huggingface#41362 | documentation | skipped | Added Hacktoberfest banner image to README.md | category not configured for this cumulative branch |
| huggingface#41333 | feature | aborted | Add DeepseekVLV2 Model | codebase moved on: DeepSeek-VL-V2 adds new model files but conflicts in multiple central auto-mapping registries whose ordering/contents have diverged; resolving would require regenerating and valida… |
| huggingface#41330 | other | skipped | Unskip and fix offline mode tests, use HF_HUB_OFFLINE, make hermetic | category not configured for this cumulative branch |
| huggingface#41329 | defect | aborted | Fix GIL=0 segfault and Add GIL=0 compat for regex paths | codebase moved on: PR wraps many slow tokenizer and conversion regex usages for Python GIL=0, but current v5 branch has replaced several tokenizers with TokenizersBackend shims and deleted deprecated… |
| huggingface#41315 | feature | aborted | [model deprecations] Define new version-based model deprecation/deletions with user warnings/exceptions | codebase moved on: deprecation infrastructure targets old PretrainedConfig and monolithic configuration_auto mappings, while current v5 branch has renamed/reworked PreTrainedConfig as a dataclass-sty… |
| huggingface#41312 | other | skipped | tests: unskip and fix offline mode test using HF_HUB_OFFLINE + hermetic cache warmup | category not configured for this cumulative branch |
| huggingface#41299 | feature | aborted | copied changes from soghomon-b:add-eval-step-limit-31561 | codebase moved on: eval-step-limit patch targets older monolithic Trainer, TrainingArguments documentation/fields, and in-file trainer tests, while current v5 branch has substantially reorganized tra… |
| huggingface#41273 | defect | aborted | fix(quantization): Skip weight initialization for quantized models | codebase moved on: quantized-loading fix conflicts across modeling_utils and many quantizer implementations; current v5 quantizer base/loading flow has different param_needs_quantization/preprocess/p… |
| huggingface#41272 | feature | aborted | feat: Add HRM Model | codebase moved on: HRM model PR is based on the older inline auto-mapping layout and docstring ignore list, while the cumulative v5 branch now imports generated auto mappings; resolving the model reg… |
| huggingface#41254 | documentation | skipped | docs: Add Hacktoberfest banner, Contributing and License sections to README | category not configured for this cumulative branch |
| huggingface#41251 | feature | aborted | Add deepseek 3.2 exp | codebase moved on: DeepSeek 3.2 experimental model conflicts with current docs toctree, FP8 integration code, package lazy imports, and generated auto configuration/model/tokenizer mappings; current … |
| huggingface#41215 | defect | aborted | Fix CLIP memory leak causing 600-800MB accumulation per batch | codebase moved on: CLIP/AIMv2/MetaCLIP2 get*_features now use the v5 BaseModelOutputWithPooling/can_return_tuple API and return projected output objects, while the PR changes these methods back to r… |
| huggingface#41202 | feature | aborted | [WIP] standardize audio kwargs | codebase moved on: audio-kwargs standardization conflicts in Whisper generation signatures/body and would require renaming input_features/audio_spectrogram through the current v5 generation flow, whi… |
| huggingface#41162 | documentation | skipped | Add Sinhala (සිංහල) translation of README | category not configured for this cumulative branch |
| huggingface#41160 | documentation | skipped | [docs] update tips syntax | category not configured for this cumulative branch |
| huggingface#41159 | feature | aborted | Support setting total_train_batch_size. | codebase moved on: total_train_batch_size patch targets older monolithic Trainer and TrainingArguments/DeepSpeed sections, while the cumulative v5 branch has reorganized optimizer/training code and s… |
| huggingface#41097 | defect | aborted | Delay and probably avoid unnecessary graph breaks in _upad_input of modeling_flash_attention_utils.py | codebase moved on: flash-attention utilities now include FA3/FA4 kernel fallback, tracing helpers, and a local indexing path that diverges substantially from the PR's older _get_unpad_data/_upad_inpu… |
| huggingface#41095 | feature | aborted | Add LLaVA-OneVision-1.5 model and related configurations | codebase moved on: LLaVA-OneVision-1.5 adds a new model but conflicts in central auto-mapping registries and check_repo ignore lists that have been reorganized/generated in the cumulative v5 branch; … |
| huggingface#41053 | feature | aborted | Qwen3 moe | codebase moved on: PR rewrites Qwen3-MoE expert/router/decoder internals with experimental env-gated grouped GEMM code, while current cumulative branch uses modular inheritance from Qwen2Moe experts/… |
| huggingface#41040 | feature | aborted | Add Keye vl 8b 1.5 | codebase moved on: new Keye VL model conflicts in generated auto-mapping registries and check_repo ignore lists, and includes fast image processor artifacts from an older tree; safe integration would… |
| huggingface#41037 | other | skipped | Tests: Apertus integration tests | category not configured for this cumulative branch |
| huggingface#41035 | documentation | skipped | docs: update speech recognition examples to use modern Common Voice d… | category not configured for this cumulative branch |
| huggingface#41024 | feature | aborted | Deprecate max_size in ConditionalDetrImageProcessor with warning | codebase moved on: PR targets older generated ConditionalDetr image processor/test layout; current branch generates the processor from modular_conditional_detr and has v5 image-processing tests cover… |
| huggingface#41022 | documentation | skipped | 🌐 [i18n-KO] Translated backbones.md to Korean | category not configured for this cumulative branch |
| huggingface#41021 | documentation | skipped | 🌐 [i18n-KO] Translated video_processors.md to Korean | category not configured for this cumulative branch |
| huggingface#41009 | feature | aborted | Add Lexa-Delta model support | codebase moved on: Lexa-Delta adds a new model against older explicit auto-mapping registries; current cumulative branch uses reorganized/generated v5 auto mappings and tokenizer mapping shapes, so r… |
| huggingface#40962 | feature | aborted | perceptron: Isaac-0.1 implementation | codebase moved on: Isaac new-model PR targets older explicit auto-mapping registries; current cumulative branch uses generated/update-based mappings and has diverged image processor mapping/check-rep… |
| huggingface#40888 | documentation | skipped | DOC Fix help for chat and serve commands | category not configured for this cumulative branch |
| huggingface#40887 | feature | aborted | Refactor output handling in generate for cleaner decoding methods | codebase moved on: generation utils now include MTP decoding, async stopping-criteria handling, flash-attention compile checks, and assisted-generation defaults from newer cumulative changes; the PRs… |
| huggingface#40877 | documentation | skipped | Bug huggingface#40833: Fix for kv_offset calculation for mixed padding | category not configured for this cumulative branch |
| huggingface#40871 | feature | aborted | Refactor benchmark utils: add type hints, GPU metrics helper, and con… | codebase moved on: PR modifies benchmark_v2/benchmark_framework.py, but that file is deleted/untracked in the current cumulative v5 branch and benchmark infrastructure has moved; applying the old ben… |
| huggingface#40870 | feature | aborted | Reduce vRAM usage during generation by allowing to transfer logits to CPU | codebase moved on: offload-logits feature patches older generation output accumulation and GenerationConfig defaults, while current cumulative branch has dataclass-style default None handling, async … |
| huggingface#45703 | other | skipped | chore(typing): add ty type checking for 10 utility files | category not configured for this cumulative branch |
| huggingface#40820 | feature | aborted | Add models to benchmarks | codebase moved on: PR targets older benchmark_v2/benchmark_framework.py plus benchmark_v2/benches files; current cumulative branch has benchmark infrastructure under benchmark_v2/framework and untrac… |
| huggingface#40738 | documentation | skipped | Docs: Clarify rjieba installation for RoFormerTokenizer | category not configured for this cumulative branch |
| huggingface#40736 | documentation | skipped | 🌐 [i18n-KO] Translated jan.md to Korean | category not configured for this cumulative branch |
| huggingface#40728 | feature | aborted | feat(serve): add OTEL | codebase moved on: PR modifies legacy src/transformers/commands/serving.py for transformers serve OTEL, but commands package is no longer tracked in this v5-era cumulative tree and untracked legacy c… |
| huggingface#40714 | documentation | skipped | Remove TF and Flax from README | category not configured for this cumulative branch |
| huggingface#40670 | feature | aborted | Add ability to run Gemma 2 models without post layer norm | codebase moved on: PR edits Gemma2/T5Gemma config init code, but current tree has strict dataclass-style configs and modular-generated model files; resolving requires porting the feature into the… |
| huggingface#40648 | other | skipped | Bump torch from 2.7.1 to 2.8.0 in /examples/flax/vision | category not configured for this cumulative branch |
| huggingface#40640 | defect | aborted | Resume training by trained samples to avoid elastic job loss or over-reading of data. | codebase moved on: Trainer training loop/state has been substantially refactored around _init_training_state/_run_epoch; PR changes older inline resume logic and large test sections, causing broad co… |
| huggingface#40637 | feature | aborted | [WIP]Add openpangu_dense model | codebase moved on: direct merge would drag unrelated upstream history/PR huggingface#40948, and cherry-picking OpenPangu model commits conflicts with current generated auto mappings/model registry layout; porti… |
| huggingface#40633 | feature | aborted | Add support for Custom Accelerate Instance in Trainer | codebase moved on: PR adds a custom Accelerator parameter to the old monolithic Trainer init and also includes broad lint/style churn; current Trainer has been refactored into a typed staged initiali… |
| huggingface#40546 | feature | aborted | Implement VibeVoice | codebase moved on: PR is a large new-model branch with 220 commits and old generated auto-mapping layout; current configuration_auto.py is generated around CONFIG_MAPPING_NAMES.update while the PR re… |
| huggingface#40524 | documentation | skipped | Use begin_of_sequence token in all sliding windows for correct model behaviour | category not configured for this cumulative branch |
| huggingface#40505 | feature | aborted | Refactor Siglip-like models | codebase moved on: SigLIP-like models have diverged in current modular/generated implementations; refactor conflicts across AimV2, Idefics2/3 and SigLIP modeling/modular files, requiring semantic reg… |
| huggingface#40493 | defect | aborted | Update dtypes to suit colab bf16 -> fp16 -> fp32. | codebase moved on: dtype resolution/loading code has been refactored into the current _get_dtype/from_pretrained flow; PR touches older modeling_utils import and dtype fallback locations and conflict… |
| huggingface#40473 | other | skipped | [tests] Unskip DeBERTaV2 tokenizer parity tests; re-enable fast/slow checks | category not configured for this cumulative branch |
| huggingface#40471 | documentation | skipped | DOC: Standardize CodeGen model card (issue huggingface#36979) | category not configured for this cumulative branch |
| huggingface#40465 | documentation | skipped | 🌐 [i18n-KO] Translated tools.md to Korean | category not configured for this cumulative branch |
| huggingface#40464 | documentation | skipped | 🌐 [i18n-KO] Translated agents.md to Korean | category not configured for this cumulative branch |
| huggingface#40448 | feature | aborted | [model] Support MiniCPM-V 4.5 | codebase moved on: MiniCPM-V 4.5 PR targets the older explicit auto mapping tables, while current v5 auto files are generated and use post-parse CONFIG/MODEL/PROCESSOR mapping updates; adding the new… |
| huggingface#40446 | feature | aborted | Add convert_segmentation_map_to_binary_masks_sorted function for hand… | codebase moved on: Mask2Former image processors have been regenerated from modular/torchvision backends and the PR adds a numpy helper/tests against the older generated slow processor; resolving safe… |
| huggingface#40425 | defect | aborted | Fix check_quantized_param method when param_value is a safetensors slice | codebase moved on: the old check_quantized_param API targeted by the PR has been replaced by param_needs_quantization and centralized loading/dtype handling; the PR conflicts across five quantizers a… |
| huggingface#40404 | documentation | skipped | update model card for gpt-j | category not configured for this cumulative branch |
| huggingface#40403 | feature | aborted | Customizable Logit Warping Strategies for Generation huggingface#40010 | codebase moved on: generation configuration and logits processing have diverged substantially (continuous batching support flags, nullable GenerationConfig defaults, new sampling controls and safety … |
| huggingface#40400 | documentation | skipped | fixed redundant words in readme.md | category not configured for this cumulative branch |
| huggingface#40395 | documentation | skipped | 🌐 [i18n-KO] Updated text_generation.md | category not configured for this cumulative branch |
| huggingface#40390 | documentation | skipped | Fix typo: 'lenght' to 'length' | category not configured for this cumulative branch |
| huggingface#40328 | feature | aborted | [rfc] Prototype to make torch.compile work with DynamicCache | codebase moved on: DynamicCache and generation compile gating have been refactored in current v5 (new cache layer constructors, cache_params lookup, ContinuousMixin/generation flow, XPU/neuron compil… |
| huggingface#40299 | feature | aborted | Remove deprecated max_size parameter from ConditionalDetr image processors | codebase moved on: DETR-family image processors have been split/regenerated into torchvision/PIL backends in current v5, while the PR removes max_size across the older generated processors and confli… |
| huggingface#40286 | feature | aborted | Add MOSS-TTSD with XY-Tokenizer | codebase moved on: adding MOSS-TTSD/XY-Tokenizer targets the older explicit auto mapping tables and repo check allowlists, while current v5 auto/model files are generated/post-processed and model add… |
| huggingface#40265 | other | skipped | Enable PLW and PLE rules | category not configured for this cumulative branch |
| huggingface#40225 | documentation | skipped | docs: clarify decoder_input_ids vs decoder_inputs_embeds usage (huggingface#39542) | category not configured for this cumulative branch |
| huggingface#40209 | feature | aborted | Add fast image processor for ViViT | codebase moved on: current v5 image processor auto mappings use backend-specific torchvision/PIL dictionaries and ViViT tests were refactored, while the PR adds an older single fast image processor f… |
| huggingface#40180 | feature | aborted | Enable native mxfp4 training support for GPT-OSS models | codebase moved on: current mxfp4 integration, quantizer, quantization config, and tests have diverged from the PR, including existing mxfp4 fixes and refactored tests; the PR adds a large autograd/ha… |
| huggingface#40177 | defect | aborted | Revert text_positions in Qwen25VL | codebase moved on: Qwen2.5-VL generation now computes packed text+vision position_ids through prepare_position_ids_for_generation and modular/generated code, whereas the PR removes an older prepare… |
| huggingface#40171 | other | skipped | Rename to CamelCase | category not configured for this cumulative branch |
| huggingface#40155 | documentation | skipped | 🌐 [i18n-KO] Translated t5.md to Korean | category not configured for this cumulative branch |
| huggingface#40149 | feature | aborted | Implemented fast image processor for VitPose | codebase moved on: fast image processor infrastructure was removed/refactored in current v5 (image_processing_utils_fast.py is deleted, auto image processor mappings are backend-specific torchvision/… |
| huggingface#40131 | documentation | skipped | add missing Arabic translations | category not configured for this cumulative branch |
| huggingface#40102 | documentation | skipped | 🌐 [i18n-KO] Translated to Korean | category not configured for this cumulative branch |
| huggingface#40064 | documentation | skipped | 🌐 [i18n-KO] Translated videomae.md to Korean | category not configured for this cumulative branch |
| huggingface#40061 | documentation | skipped | 🌐 [i18n-KO] Translated vitdet.md to Korean | category not configured for this cumulative branch |
| huggingface#40047 | documentation | skipped | Update wavlm.md to match new model card template | category not configured for this cumulative branch |
| huggingface#39987 | feature | aborted | Add a VGGT(Visual Geometry Grounded Transformer) model compatible with huggingface transfromers | codebase moved on: PR adds VGGT against older static auto-mapping files, while current branch uses generated/updated auto registry structure; direct merge conflicts in model registries and resolving … |
| huggingface#39962 | defect | aborted | Use torch._check instead of a test to make the model Gemma3 exportable | codebase moved on: PR touches many multimodal modeling and tests to replace shape assertions with torch._check, but current branch has substantially refactored generated/modular model files and the d… |
| huggingface#39931 | feature | aborted | Registers StaticCache serialization functions for torch.export.export | codebase moved on: StaticCache and executorch integration have changed since the PR; direct merge conflicts in cache_utils, integrations/executorch.py, and cache tests. Current cache code already con… |
| huggingface#39922 | documentation | skipped | 🌐 [i18n-KO] Translated attention_interface.md to Korean | category not configured for this cumulative branch |
| huggingface#39920 | documentation | skipped | 🌐 [i18n-KO] Updated ko/perf_train_special.md | category not configured for this cumulative branch |
| huggingface#39917 | documentation | skipped | 🌐 [i18n-KO] Updated ko/perf_train_cpu.md | category not configured for this cumulative branch |
| huggingface#39901 | documentation | skipped | 🌐 [i18n-KO] Translated fp_quant to Korean | category not configured for this cumulative branch |
| huggingface#39899 | feature | aborted | [model] Support MiniCPM-V 4.0 | codebase moved on: PR adds MiniCPM-V 4.0 against older static auto registries; direct merge conflicts in all auto mapping files and the model package registry, and the PR also carries obsolete genera… |
| huggingface#39886 | documentation | skipped | 🌐 [i18n-KO] Translated perf_train_gaudi.md to Korean | category not configured for this cumulative branch |
| huggingface#39859 | feature | aborted | WIP: Initial support for bnb 4bit on any nn.Parameter | codebase moved on: PR implements bnb target_parameters through older create_quantized_param/check_quantized_param paths, while current branch has refactored 4-bit loading into param_needs_quantizatio… |
| huggingface#39831 | other | skipped | refactor(modeling_llama): make RotaryEmbedding default path explicit | category not configured for this cumulative branch |
| huggingface#39807 | documentation | skipped | 🌐 [i18n-KO] Translated bamba.md to Korean | category not configured for this cumulative branch |
| huggingface#39796 | feature | aborted | [pipelines] text-to-audio pipeline standardization | codebase moved on: text-to-audio standardization touches pipeline behavior plus CSM/Dia/Qwen2.5-Omni model internals and broad tests; current branch has diverged in generation/modeling/pipeline/test … |
| huggingface#39792 | defect | aborted | Served models handle with nested content | codebase moved on: PR targets the legacy transformers serve command/test paths, but in the current tracked tree src/transformers/commands/serving.py and tests/commands/test_serving.py are not tracked… |
| huggingface#39772 | defect | aborted | Fix missing initializations for models created in 2022 | codebase moved on: PR tries to add missing initialization tests and _init_weights assignments across 2022-era models, but current cumulative branch has diverged in almost every touched model/test fil… |
| huggingface#39760 | feature | aborted | [Draft] Add Llasa TTS family of models | codebase moved on: Llasa is a draft new-model addition against older auto-mapping/check_repo and docs TOC layout. Direct merge conflicts in auto configuration registration, generated docs TOC, and re… |
| huggingface#39756 | defect | aborted | Fix rope_deltas corruption in Qwen2.5VL during CFG generation | codebase moved on: PR targets older Qwen2/Qwen2.5-VL and GLM4V generation/rope_deltas code, while current branch has a redesigned 3D position-id preparation and expanded multimodal generation helpers… |
| huggingface#39751 | documentation | skipped | 🌐 [i18n-KO] Translated text-to-speech.md to Korean | category not configured for this cumulative branch |
| huggingface#39722 | feature | aborted | [Feat] Adding Intern-S1 | codebase moved on: Intern-S1 is a broad new-model addition against older auto mappings, processor/video processor registration, docs TOC, and check_repo expectations. Current branch conflicts in all … |
| huggingface#39718 | documentation | skipped | Fix SigLIP2 documentation model/processor mismatch | category not configured for this cumulative branch |
| huggingface#39708 | documentation | skipped | 🌐[i18n-bn] Introduce Bengali version of Transformers documentation | category not configured for this cumulative branch |
| huggingface#39632 | documentation | skipped | fix dead NVIDIA link | category not configured for this cumulative branch |
| huggingface#39631 | feature | aborted | [serve] Add speech-to-text | codebase moved on: the serving command/docs files targeted by the PR are currently untracked in this cumulative v5-era worktree and would be overwritten by a direct merge; the PR is a large 47-commit… |
| huggingface#39588 | feature | validation_failed | WIP, reference modeling | merge was clean, but PR-local ruff found many new style errors plus an undefined MultiModalProjector symbol in the added reference_vlm model |
| huggingface#39575 | documentation | skipped | 🌐 [i18n-KO] Translated vitpose.md to Korean | category not configured for this cumulative branch |
| huggingface#39563 | documentation | skipped | 🌐 [i18n-KO] Translated vision-encoder-decoder.md to Korean | category not configured for this cumulative branch |
| huggingface#39559 | documentation | skipped | 🌐 [i18n-KO] Translated main_classes/deepspeed.md to Korean | category not configured for this cumulative branch |

evalstate · 2026-04-29T19:46:15Z

Feature + defect flow status table (part 4/4)

evalstate and others added 30 commits April 29, 2026 06:17

Merge branch 'mergeability-pr-45075' into features-and-defects-750

8405276

# Conflicts: # src/transformers/conversion_mapping.py

Merge branch 'mergeability-pr-45040' into features-and-defects-750

773209e

Merge branch 'mergeability-pr-44989' into features-and-defects-750

358420e

Fix train_batch_size and eval_batch_size to respect split_batches config

936f92c

Merge branch 'mergeability-pr-44952' into features-and-defects-750

6377e75

Merge branch 'mergeability-pr-44940' into features-and-defects-750

9da1c7f

# Conflicts: # src/transformers/initialization.py

Merge branch 'mergeability-pr-44923' into features-and-defects-750

b2da860

# Conflicts: # src/transformers/tokenization_utils_tokenizers.py

Merge branch 'mergeability-pr-44907' into features-and-defects-750

374660d

# Conflicts: # src/transformers/models/colqwen2/modeling_colqwen2.py # src/transformers/models/colqwen2/modular_colqwen2.py # src/transformers/models/paligemma/modeling_paligemma.py

Merge branch 'mergeability-pr-44893' into features-and-defects-750

309c85e

Merge branch 'mergeability-pr-44891' into features-and-defects-750

02e557f

Merge branch 'mergeability-pr-44889' into features-and-defects-750

2950cb8

Merge branch 'new-model-original-code' into granite_speech_plus

552fba0

Merge branch 'mergeability-pr-45694' into features-and-defects-750

9019aa5

Merge branch 'mergeability-pr-44836' into features-and-defects-750

cb6444a

Merge branch 'mergeability-pr-44827' into features-and-defects-750

90d49d3

Merge branch 'mergeability-pr-44793' into features-and-defects-750

a28192f

Merge branch 'mergeability-pr-44781' into features-and-defects-750

d534dbe

# Conflicts: # tests/test_tokenization_common.py

Merge branch 'mergeability-pr-44771' into features-and-defects-750

e40fdff

# Conflicts: # src/transformers/models/pi0/image_processing_pi0.py

add compressed-tensor fp8 integeration

5f46f04

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

Merge branch 'mergeability-pr-44731' into features-and-defects-750

297b062

Merge branch 'mergeability-pr-44724' into features-and-defects-750

4102d18

# Conflicts: # src/transformers/models/auto/configuration_auto.py # src/transformers/models/auto/image_processing_auto.py # src/transformers/models/auto/modeling_auto.py

Merge branch 'mergeability-pr-44713' into features-and-defects-750

74e7ea5

# Conflicts: # src/transformers/models/colqwen2/modeling_colqwen2.py # src/transformers/models/colqwen2/modular_colqwen2.py

Merge branch 'mergeability-pr-44697' into features-and-defects-750

8e8a6bd

# Conflicts: # src/transformers/integrations/flex_attention.py # src/transformers/utils/generic.py

Rerun modular_model_converter

27be42a

Merge branch 'mergeability-pr-44680' into features-and-defects-750

d6cb5cf

Merge branch 'mergeability-pr-44676' into features-and-defects-750

bf523b9

Merge branch 'mergeability-pr-45695' into features-and-defects-750

998596f

update

f5a3168

Signed-off-by: jiqing-feng <jiqing.feng@intel.com>

Merge branch 'mergeability-pr-44664' into features-and-defects-750

25ab701

# Conflicts: # src/transformers/models/qwen3_5/modeling_qwen3_5.py

Merge branch 'mergeability-pr-44662' into features-and-defects-750

5c1770e

# Conflicts: # src/transformers/conversion_mapping.py # src/transformers/models/auto/configuration_auto.py # src/transformers/models/auto/image_processing_auto.py # utils/check_repo.py

evalstate added 23 commits April 29, 2026 19:46

Apply PR huggingface#39698: fix Exaone4 sliding window layer types

9c3cc50

Merge branch 'mergeability-pr-39697' into features-and-defects-750

8f73042

Apply PR huggingface#39690: allow custom hf_quantizer

2e2779e

Apply PR huggingface#39683: respect disabled torch dynamo

952a5f6

Apply PR huggingface#39674: scale loss by data parallel size

4d1d1fc

Apply PR huggingface#39599: tolerate missing trainer state

39c0527

Apply PR huggingface#39560: save best checkpoints on eval

864cf5e

Apply PR huggingface#39493: pin mistral-common extras

fcf8baf

Apply PR huggingface#39491: skip quantized weight init

8c95ec9

Apply PR huggingface#39468: skip bnb dispatch

7b6ed57

Apply PR huggingface#39435: add Bart mask regression

1d68999

Merge branch 'mergeability-pr-39309' into features-and-defects-750

8205757

# Conflicts: # src/transformers/pipelines/audio_classification.py # src/transformers/pipelines/automatic_speech_recognition.py # tests/pipelines/test_pipelines_automatic_speech_recognition.py

Merge branch 'mergeability-pr-39257' into features-and-defects-750

8d8de6f

# Conflicts: # src/transformers/utils/generic.py

Port Qwen3 MoE empty router logits fix (huggingface#39206)

74def78

Port chat extra dependency group (huggingface#39183)

d2dfb4c

Port MoE fullgraph compile disable (huggingface#39108)

f5c57c0

Apply PR huggingface#39103: fix Gemma3n audio config naming

bb2f6c4

Applied from upstream PR huggingface#39103 (1d85c39) with local patch because the direct PR merge conflicts with current Gemma3n v5/strict config docstring structure.

Merge branch 'mergeability-pr-39047' into features-and-defects-750

618429c

# Conflicts: # src/transformers/loss/loss_utils.py

Merge branch 'mergeability-pr-39037' into features-and-defects-750

32b7952

# Conflicts: # src/transformers/models/kosmos2/modeling_kosmos2.py

Merge branch 'mergeability-pr-39009' into features-and-defects-750

33afdec

Merge branch 'mergeability-pr-38888' into features-and-defects-750

27c3403

Merge branch 'mergeability-pr-38886' into features-and-defects-750

9ffa016

# Conflicts: # src/transformers/quantizers/quantizer_bnb_4bit.py # src/transformers/quantizers/quantizer_bnb_8bit.py

Merge branch 'mergeability-pr-38884' into features-and-defects-750

199afea

# Conflicts: # src/transformers/models/llama4/convert_llama4_weights_to_hf.py

evalstate changed the title ~~Cumulative feature and defect fixes from recent Transformers PRs~~ Cumulative feature and defect updates from recent Transformers PRs Apr 29, 2026

Add PR classifications artifact for feature-defect run

85172b4

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Cumulative feature and defect updates from recent Transformers PRs#42

Cumulative feature and defect updates from recent Transformers PRs#42
evalstate wants to merge 1808 commits intomainfrom
features-and-defects-750

evalstate commented Apr 29, 2026 •

edited

Loading

Uh oh!

evalstate commented Apr 29, 2026

Uh oh!

evalstate commented Apr 29, 2026

Uh oh!

evalstate commented Apr 29, 2026

Uh oh!

evalstate commented Apr 29, 2026

Uh oh!

evalstate commented Apr 29, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

Conversation

evalstate commented Apr 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Cumulative feature + defect PR mergeability branch

Status counts

Category counts

Validation note

Uh oh!

evalstate commented Apr 29, 2026

Generated mergeability report

All defects cumulative merge report

Run context

Record counts

Cumulative classifications

Cumulative merge states

Baseline validation for this turn

PRs processed this turn

Newly merged/applied PRs this turn

Validation results this turn

Recent merged/applied PRs

Recent aborted PRs

Resume state

Uh oh!

evalstate commented Apr 29, 2026

Feature + defect flow status

Status counts

Category counts

Merged / applied / already-present records (393)

Uh oh!

evalstate commented Apr 29, 2026

Feature + defect flow status table (part 2/4)

Rejected / not included records (379)

Uh oh!

evalstate commented Apr 29, 2026

Feature + defect flow status table (part 3/4)

Uh oh!

evalstate commented Apr 29, 2026

Feature + defect flow status table (part 4/4)

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

evalstate commented Apr 29, 2026 •

edited

Loading