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

1. 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

Feature + defect flow status

Processed terminal records: 772

Status counts

Category counts

Merged / applied / already-present records (393)

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

| huggingface#42098 | defect | merged | merge | unavailable | Fix mel length computation in Qwen2-Audio | merged with a local docstring conflict resolution; mel-length fix applied; compileall and style passed; light validation unavailable due baseline untracked-fil… |
| huggingface#42051 | defect | merged | merge | unavailable | Fix model_input_names singleton issue causing shared state | merge clean; compileall and style passed; light validation unavailable due baseline untracked-file checkout issue |
| huggingface#41973 | defect | applied | patch | unavailable | Fix import error with huggingface_hub v1.0.0+ | direct merge conflicted with current lazy hub-import refactor; applied the two remaining import-location fixes only; compileall and style passed; light validat… |
| huggingface#41928 | defect | applied | patch | unavailable | fix: add clear error message when mistral-common is missing for AutoTokenizer loading Voxtral | direct merge conflicted with current AutoTokenizer backend refactor; applied the Voxtral missing-mistral-common error check to the current code; compileall and… |
| huggingface#41904 | defect | applied | patch | unavailable | Fix inaccurate eval and train loss computation with variable batch sizes | direct merge conflicted in trainer.py; applied variable-batch loss weighting to current examples and Trainer loop; compileall and style passed; light validatio… |
| huggingface#41901 | feature | merged | merge | unavailable | [executorch] Update pytree registration for DynamicCache | merge clean; compileall and style passed; light validation unavailable because tests_fetcher cannot run with existing untracked generated files and baseline li… |
| huggingface#41895 | feature | merged | merge | unavailable | Add Telugu Sentiment Classification Example using DistilBERT | merge clean; compileall and style passed; light validation unavailable because tests_fetcher cannot checkout merge parent over existing untracked generated fil… |
| huggingface#41879 | defect | already_present | none | not_run | Fix/processor multiple tokenizers | current ProcessorMixin already excludes all tokenizer attributes from to_dict, saves non-primary tokenizers in attribute subfolders, and loads additional token… |
| huggingface#41855 | defect | merged | merge | unavailable | Add Mistral tokenizer missing methods | merge had a docstring conflict resolved in current tokenizer wrapper; formatter-only change amended; compileall and style passed; light validation unavailable … |
| huggingface#41851 | defect | already_present | none | not_run | Fix deepcopy in ProcessorMixin.to_dict for GemmaTokenizerFast | same ProcessorMixin multi-tokenizer/deepcopy fix as PR 41879 is already covered by the current generalized implementation; no code change needed |
| huggingface#41844 | defect | applied | patch | unavailable | Fix FSDPv2 checkpoint saving on TPU by using recursive unwrap | direct PR-head merge conflicted in trainer.py after TPU checkpoint saving moved to integrations/tpu.py; applied recursive FSDPv2 unwrap to the current helper; … |
| huggingface#41827 | defect | merged | merge | unavailable | [Flash Attention] Disable packed sequences with pos ids only during torch compile | merged with local conflict resolution in current flash attention utilities; compileall and style passed; light validation unavailable/timed out after creating … |
| huggingface#41798 | feature | merged | merge | unavailable | p-less Sampling: A Robust Hyperparameter-Free Approach for LLM Decoding | merged with a local GenerationConfig conflict resolution preserving current None defaults and warning gating; compileall and style passed; light validation una… |
| huggingface#41797 | feature | already_present | none | not_run | Add deepseek ocr | current cumulative branch already contains DeepSeek-OCR model/docs/registration; direct merge only exposed conflicts in generated auto mappings and repo-check … |
| huggingface#41776 | feature | merged | merge | unavailable | Add safety checking infrastructure for text generation | merged with a local docs conflict resolution; narrow ruff fixes on PR files amended; compileall and style passed; light validation unavailable/timed out |
| huggingface#41734 | defect | merged | merge | unavailable | Fix CUDA errors in sharded generation with Qwen3 | merged with local generation-utils conflict resolution and narrow ruff fix; compileall and style passed; light validation unavailable/timed out |
| huggingface#41724 | defect | already_present | none | not_run | Fix confusing warning in EncoderDecoderModel when creating decoder_input_ids from labels | intended warning split is already present in current cumulative branch; attempted PR merge only conflicted on an unrelated decoder_attention_mask expression |
| huggingface#41718 | defect | merged | merge | unavailable | AutoTokenizer: clear ImportError when loading Voxtral without mistral-common + unit test | merged with local tokenization_auto conflict resolution preserving v5 remote-code guard; narrow ruff fix amended; compileall and style passed; light validation… |
| huggingface#41701 | defect | merged | merge | unavailable | Fix qwen3_vl mix precision dtype | merged cleanly; compileall and style passed after removing transient untracked files left by light validation; light validation unavailable from tests_fetcher … |
| huggingface#41698 | defect | applied | cherry-pick | unavailable | Fix tokenizer check script: safe dataset access, default checkpoints, and tested in dry-run mode | cherry-picked PR fix commit with narrow conflict resolution and ruff amend; compileall and style passed; light validation unavailable from tests_fetcher as in … |
| huggingface#41687 | defect | merged | merge | unavailable | fix(data): Handle integer labels in DataCollatorWithFlattening | merged with local data_collator/test conflicts resolved for current pack_sequence_labels implementation; narrow ruff format amended; compileall and style passe… |
| huggingface#41631 | defect | already_present | none | not_run | Incorrect access of dataset field fixed | same dataset field access fix is already included via PR 41698 applied earlier in this cumulative branch |
| huggingface#41606 | defect | already_present | none | not_run | fix(processing): Filter kwargs in ProcessorMixin call to prevent Type… | ProcessorMixin has since been refactored to split merged kwargs by modality and pass only images/videos/audio/text kwargs to each subprocessor, so the PR fix i… |
| huggingface#41594 | feature | merged | merge | unavailable | Add beginner-friendly sentiment analysis example | merged cleanly; ruff-only fixes on new example files amended; compileall and style passed; light validation unavailable from tests_fetcher timeout as in baseli… |
| huggingface#41593 | feature | merged | merge | unavailable | examples: add multi-label text classification (BCEWithLogitsLoss, met… | merged with README conflict resolved by preserving existing text-classification README plus adding multi-label section; ruff-only fixes amended; compileall and… |
| huggingface#41592 | defect | already_present | none | not_run | Improve AutoTokenizer error message for Voxtral models missing mistral-common | current AutoTokenizer already raises a clear ImportError for model_type == 'voxtral' when mistral-common is unavailable, so the intended error-message fix is p… |
| huggingface#41561 | feature | merged | merge | unavailable | Optimize Mamba2 memory usage by replacing broadcast with einsum | merged cleanly; compileall and style passed after removing validation-created untracked files; light validation unavailable from tests_fetcher timeout as in ba… |
| huggingface#41528 | feature | already_present | none | not_run | Add position encoding interpolation to DeiT | current DeiT model already has interpolate_pos_encoding support, including dynamic-shape-safe position interpolation and slow inference coverage; the stale PR … |
| huggingface#41524 | feature | applied | patch | unavailable | Add max_eval_batches argument to TrainingArguments | applied max_eval_batches to current TrainingArguments and evaluation_loop; compileall and style passed after removing stale untracked validation artifacts; lig… |
| huggingface#41523 | feature | already_present | none | not_run | Add test coverage for ConvNextImageProcessorFast | current v5 image-processing test mixin already discovers ConvNext pil/torchvision processors from IMAGE_PROCESSOR_MAPPING_NAMES, so fast/backend coverage is pr… |
| huggingface#41522 | defect | applied | patch | unavailable | Fix _init_weights to safely skip int8 quantized weights | applied a non-floating weight guard before Qwen2.5-VL initialization while preserving current rotary initialization; compileall and style passed; light validat… |
| huggingface#41521 | defect | merged | merge | unavailable | Fix forced_bos_token_id not set in generation_config | merged cleanly; compileall and style passed; light validation unavailable from tests_fetcher timeout as in baseline |
| huggingface#41490 | defect | already_present | none | not_run | Fix _init_weights to safely skip int8 tensors in Qwen2_5_VL model | same intended Qwen2.5-VL non-floating/int8 initialization fix was already applied from overlapping PR huggingface#41522 in this cumulative branch |
| huggingface#45702 | defect | merged | merge | unavailable | Reorder decorators for autodoc and dataclass | merged with a local conflict resolution in qwen3_5 preserving existing MTP code while applying the decorator-order fix; compileall and style passed; light vali… |
| huggingface#45699 | feature | already_present | none | not_run | Add FP8 kernel acceleration for compressed-tensors quantized models | PR head changes are already reachable in the cumulative branch, so no additional code change was needed |
| huggingface#45694 | defect | already_present | none | not_run | Fix train_batch_size and eval_batch_size to respect split_batches config | PR head changes are already reachable in the cumulative branch, so no additional code change was needed |
| huggingface#45690 | feature | already_present | none | not_run | [serve] Support for reasoning | PR head changes are already reachable in the cumulative branch, so no additional code change was needed |
| huggingface#45687 | defect | already_present | none | not_run | fix: Made histc_input robust for broader hardware | PR head changes are already reachable in the cumulative branch, so no additional code change was needed |
| huggingface#45683 | defect | already_present | none | not_run | Exclude audio modules from conversion process | PR head changes are already reachable in the cumulative branch, so no additional code change was needed |
| huggingface#45682 | defect | already_present | none | not_run | FIX Restore LoRA hotswapping functionality | PR head changes are already reachable in the cumulative branch, so no additional code change was needed |
| huggingface#45681 | defect | already_present | none | not_run | Restore TokenizersBackend override for DeepSeek V3/R1 tokenizer dispatch | current cumulative branch already honors TokenizersBackend/PythonBackend overrides before trusting tokenizer_config_class and includes DeepSeek regression cove… |
| huggingface#45678 | defect | already_present | none | not_run | Fix shared config mutation issue in flash_attn_from_config | PR head changes are already reachable in the cumulative branch, so no additional code change was needed |
| huggingface#41458 | feature | merged | merge | unavailable | Adding ScatterMoE kernel support for Granite models. | merge clean; compileall and style passed; validation unavailable because light validation tests_fetcher timed out as in baseline |
| huggingface#41441 | feature | merged | merge | unavailable | Enhance the handling of Union types in HfArgumentParser | merge clean; compileall and style passed; validation unavailable because light validation tests_fetcher timed out as in baseline |
| huggingface#41356 | feature | already_present | none | not_run | Add DEIMv2 model, image processor, and basic tests | DEIMv2 model/docs/tests are already present on the cumulative branch; direct merge conflicts add/add with the existing generated/modular implementation, so no … |
| huggingface#41349 | feature | merged | merge | unavailable | Create (3d_parrallel_v2.py) - Add 3D parallelism training example script | merge clean; amended ruff-only fixes in changed example file; compileall and style passed; validation unavailable because light validation tests_fetcher timed … |
| huggingface#41319 | defect | already_present | none | not_run | Use torch._check instead of a test in Gemma3Model | current Gemma3 image-token check already uses torch_compilable_check instead of a data-dependent tensor indexing assertion, preserving torch.export compatibili… |
| huggingface#41313 | defect | merged | merge | unavailable | Jitter noise PR | merged with a local conflict resolution preserving current classifier dtype cast while applying jitter noise only to routing_states; compileall and style passe… |
| huggingface#41304 | defect | merged | merge | unavailable | Fix equality-vs-assignment bug in GptqHfQuantizer.update_device_map | merged with a local conflict resolution applying the device_map assignment fix to the current GPTQ quantizer; compileall and style passed; light validation una… |
| huggingface#41291 | feature | already_present | none | not_run | Add-Deimv2 | DEIMv2 integration is already present on the cumulative branch with newer modular/generated files; direct merge only conflicts add/add with existing DEIMv2 doc… |
| huggingface#41239 | defect | applied | patch | unavailable | Add num_hidden_layers to t5gemma's top level config | applied the T5Gemma top-level num_hidden_layers fix to current modular/generated config after direct merge conflicted with the newer strict dataclass __post_in… |
| huggingface#41224 | feature | merged | merge | unavailable | Add DINOv3ViTForImageClassification support | merged with a small import conflict resolution in DINOv3ViT modeling/modular files; compileall and style passed; light validation unavailable from tests_fetche… |
| huggingface#41169 | defect | merged | merge | unavailable | Fix TorchDynamo crash in StaticCache by validating offloading and offload_only_non_sliding arguments | merged cleanly; compileall and style passed; light validation unavailable from tests_fetcher timeout as in baseline |
| huggingface#41144 | feature | merged | merge | unavailable | Support automatic conversion from zero checkpoint to universal checkpoint. | merged cleanly; amended narrow current-code fixes for self.args and bool annotation; compileall and style passed after removing untracked stale validation-bloc… |
| huggingface#41132 | defect | applied | patch | unavailable | fix(SpeechT5Config): missing annotation on inputs_to_logits_ratio property | direct merge conflicted with old SpeechT5Config init layout; applied the one-line @Property fix to current config; compileall and style passed; light validatio… |
| huggingface#41121 | defect | already_present | none | not_run | fix: resolve the unexpected video frame drop issue of the InternVL model with multiple video inputs | current InternVL processor already uses per-video patch boundary indices and has a frames-binding test, so the PR's multi-video frame-drop fix is present; dire… |
| huggingface#41116 | feature | already_present | none | not_run | Add MiniCPM3 | MiniCPM3 support is already present on the cumulative branch from the newer overlapping PR huggingface#45613, including config/model/modular files and auto mappings; no a… |
| huggingface#41105 | defect | applied | patch | unavailable | Fix is_torch_neuroncore_available | applied the narrow NeuronCore check_device fix to current import_utils because PR head carries older import_utils history; compileall and style passed; light v… |
| huggingface#41077 | defect | already_present | none | not_run | Fix: add num_hidden_layers property to T5GemmaConfig and add test for use_cache | same T5Gemma num_hidden_layers fix was already applied from overlapping PR huggingface#41239 to both modular and generated config; no code change needed |
| huggingface#41075 | defect | merged | merge | unavailable | Fix Qwen3 deterministic generation when do_sample=False and num_beams=1 for Greedy Decoding | merged cleanly; compileall and style passed; light validation unavailable from tests_fetcher timeout as in baseline |
| huggingface#41041 | feature | merged | merge | unavailable | [WIP] Add YuE model | merged cleanly; amended narrow ruff fixes in new YuE files; compileall and style passed; light validation unavailable from tests_fetcher timeout as in baseline |
| huggingface#41033 | feature | merged | merge | unavailable | feat: make audio feature extractors torch.export-able | merged with a trivial tests/test_executorch.py import conflict resolved; compileall and style passed; light validation unavailable from tests_fetcher timeout a… |
| huggingface#40976 | feature | merged | merge | unavailable | Better defaults for assisted generation | merged with current candidate-generator min/max length handling kept; amended narrow UP045 ruff fix; compileall and style passed; light validation unavailable … |
| huggingface#40954 | defect | already_present | none | not_run | Fix Issue huggingface#40913: Respect user-provided chat_template parameter in processor creation | current ProcessorMixin already preserves user-provided chat_template because processor_dict is updated with kwargs after loading chat_templates; PR test file i… |
| huggingface#40908 | defect | merged | merge | unavailable | Fix incompatible with | merged cleanly; compileall and style passed; light validation unavailable as in baseline |
| huggingface#40898 | feature | merged | merge | unavailable | Adding [T5/MT5/UMT5]EncoderForSequenceClassification | merged with a trivial T5 test import conflict resolved; amended narrow py313 typing fixes in changed model files; compileall and changed-file ruff passed; full… |
| huggingface#40861 | feature | merged | merge | unavailable | Support n_groups>1 for mamba2 | merged with a narrow Mamba2 init conflict resolution preserving current lazy weight initialization; compileall and changed-file ruff passed; full style and lig… |
| huggingface#40857 | defect | already_present | none | not_run | Token | current Trainer already computes train_tokens_per_second from current-session tokens via _initial_num_input_tokens_seen; direct merge conflicts against the ref… |
| huggingface#40840 | feature | merged | merge | unavailable | feat: add qwen2 pruning support | merged with a small Qwen2 conflict resolved against current layer_type handling and single-quote style normalized; compileall and changed-file ruff passed; ful… |
| huggingface#40790 | defect | applied | patch | unavailable | Handle loading non-existent checkpoints or corrupted checkpoints. | applied minimal missing/corrupt checkpoint handling in Trainer and get_last_checkpoint; compileall and changed-file ruff passed; full style/light validation un… |
| huggingface#40783 | defect | merged | merge | unavailable | Fix None quantization_config equivalence with omitted param in AutoModel.from_pretrained | merged cleanly; compileall and changed-file ruff passed; full style/light validation unavailable as in baseline |
| huggingface#40759 | feature | already_present | none | not_run | feat: add qwen3 pruning support | Qwen3 structured-pruning changes are already contained in the earlier merged PR huggingface#40840; merge reported already up to date |
| huggingface#40756 | feature | merged | merge | unavailable | [WIP] Add Canary | merged cleanly; compileall and changed-file ruff passed; full style/light validation unavailable as in baseline |
| huggingface#40755 | feature | merged | merge | unavailable | [TimesFM] Add support for forecasting with covariates | merged cleanly; compileall and changed Python-file ruff passed; full style/light validation unavailable as in baseline |
| huggingface#40740 | defect | merged | merge | unavailable | Configure assistant model's generation_config with user parameters | merge clean; validation unavailable due pre-existing ruff failures and untracked files blocking tests_fetcher checkout |
| huggingface#40695 | defect | already_present | none | not_run | remove vision2seq vs image-text-to-text | intended fix already present: current modeling_auto.py has no MODEL_FOR_VISION_2_SEQ mapping containing Qwen vision2seq entries |
| huggingface#40587 | feature | merged | merge | unavailable | feat(utils): add vision utils for embedding images and getting the hidden size | merge clean; validation unavailable because baseline already fails ruff on pre-existing untracked legacy/restored files and light validation tests_fetcher time… |
| huggingface#40563 | defect | already_present | none | not_run | fix to get output of intermediate output of dinov3 for more use case | intended DINOv3 intermediate hidden-state support is already present: current DINOv3ViTModel delegates to DINOv3ViTEncoder with capture_outputs and returns out… |
| huggingface#40520 | feature | merged | merge | unavailable | [generate] add faster stop_strings stopping criteria | merge completed with a small test conflict resolved; validation unavailable because baseline already fails ruff on pre-existing untracked/restored files and li… |
| huggingface#40515 | feature | merged | merge | unavailable | Add Context-Aware Tokenizer Selection Utility Based on Corpus Analysis | merge clean; validation unavailable because baseline already fails ruff on pre-existing untracked/restored files and light validation tests_fetcher times out; … |
| huggingface#40492 | defect | merged | merge | unavailable | avoid divid zero errors. | merge completed; masking_utils None guard was already present in current code, and image_processing_utils/debug_utils changes were applied; validation unavaila… |
| huggingface#40438 | defect | merged | merge | unavailable | Resolve automatic label name detection when single label provided | merge completed with Trainer conflict resolved by applying the label_names=[label] reset to the current staged Trainer initialization; compileall and PR-local … |
| huggingface#40392 | defect | applied | cherry-pick | unavailable | Remove debug print statement from ShieldGemma2 conversion script | cherry-pick applied the one-line debug-print removal without dragging older branch history; compileall passed, but full validation unavailable due pre-existing… |
| huggingface#40388 | feature | applied | cherry-pick | unavailable | Add fromjson filter to Jinja2 chat templates | cherry-pick applied the fromjson filter and tests without pulling older branch history; compileall and PR-local ruff passed, full validation unavailable due pr… |
| huggingface#40385 | defect | applied | cherry-pick | unavailable | Fix typo: 'seperate' -> 'separate' in mm_grounding_dino conversion sc… | cherry-pick applied the MM Grounding DINO conversion key correction; compileall and PR-local ruff passed, full validation unavailable due pre-existing baseline… |
| huggingface#40358 | defect | applied | patch | unavailable | Fix MXFP4 mlp_forward to handle 2D and 3D hidden_states shapes for multi-turn chat | applied the MXFP4 mlp_forward dimensionality fix only; direct merge/cherry-pick would include older-history test/CI churn, while current integration code has d… |
| huggingface#40244 | defect | merged | merge | unavailable | add-loftr-keypoints-to-map | merge required a trivial auto-mapping conflict resolution; validation unavailable due baseline ruff failures and tests_fetcher backend error |
| huggingface#40221 | defect | applied | patch | unavailable | FIX: enable load_best_model_at_end within SaveStrategy.BEST and initialize metric_for_best_model as loss when SaveStrategy.BEST | direct merge conflicted in trainer files after test/module refactors; ported remaining metric_for_best_model default change, while trainer SaveStrategy.BEST ch… |
| huggingface#40208 | defect | applied | patch | unavailable | Save only model sharded sd | direct merge conflicted in trainer after refactors; ported the output_dir creation, FSDP sharded model save path, and removal of the incompatible validation; v… |
| huggingface#40148 | defect | merged | merge | unavailable | Update utils.py: fix nan | merge clean; compileall and PR-local ruff passed, but full validation unavailable due pre-existing baseline ruff failures and light-validation tests_fetcher ch… |
| huggingface#40123 | defect | already_present | none | unavailable | Lazily import torchao Int4WeightOnlyConfig to avoid side effects | intended torchao eager-import side effect is already absent in current code after quantization/loading refactors; direct merge conflicts in obsolete modeling_u… |
| huggingface#40115 | feature | already_present | none | unavailable | [layer_types] update layer_types with conv | conv is already allowed in current ALLOWED_LAYER_TYPES and Lfm2Config already derives conv/full_attention layer_types in the refactored dataclass config layout… |
| huggingface#40114 | defect | applied | patch | unavailable | Fix torch.export compatibility for Mixtral MoE models | ported the Mixtral inference-time static expert loop to current optimized modular/generated MixtralExperts; compileall and PR-local ruff passed, but full valid… |
| huggingface#40092 | feature | merged | merge | unavailable | Optimize LlamaAttention by fusing QKV projections | merge clean; amended a PR-local blank-line whitespace fix; compileall and PR-local ruff passed, but full validation unavailable due baseline ruff failures and … |
| huggingface#40090 | defect | applied | patch | unavailable | Fix RuntimeError when loading quantized models with int8 weights (huggingface#39366) | ported non-floating weight/bias initialization guards to current modeling_utils; compileall and PR-local ruff passed, but full validation unavailable due basel… |
| huggingface#40065 | defect | applied | patch | unavailable | Delay float32 upcast in ForCausalLMLoss after filtering ignore_index | ported the ForCausalLMLoss ignore_index filtering before float32 upcast; compileall and PR-local ruff passed, but full validation unavailable due baseline ruff… |
| huggingface#40059 | defect | applied | patch | unavailable | Fix Inefficient GELU implementation in GPT2 | ported GPT-2 default activation from gelu_new to fused gelu; compileall and PR-local ruff passed, but full validation unavailable due baseline ruff failures an… |
| huggingface#40058 | feature | applied | patch | unavailable | GGUF Qwen2VL | ported Qwen2VL GGUF config/model-type/tokenizer converter support and coverage to current GGUF files; compileall and PR-local ruff passed, but full validation … |
| huggingface#40055 | feature | merged | merge | unavailable | Auto-log parallelism info to wandb.config using HF Accelerate | merge clean; amended PR-local W293/S110 lint fixes in wandb callback; compileall and PR-local ruff passed, but full validation unavailable due baseline ruff fa… |
| huggingface#40023 | feature | already_present | none | unavailable | Add support for SDPA for OWLViT and OWLv2 | OWLViT/OWLv2 already use ALL_ATTENTION_FUNCTIONS with _supports_sdpa and have SDPA parity tests in current branch; direct merge conflicted only with the newer … |
| huggingface#40022 | defect | applied | patch | unavailable | fix: resolve dropout type error in DogeDecoder | ported Doge MoE tuple unwrapping before dropout plus router_gate zero initialization to current modular/generated Doge; compileall and PR-local ruff passed, bu… |
| huggingface#39999 | defect | applied | patch | unavailable | allow TP to work in ND-parallel with fsdp cpu ram efficient loading | ported TP+FSDP meta device_map handling into current initialize_tensor_parallelism helper; compileall and PR-local ruff passed, but full validation unavailable… |
| huggingface#39997 | defect | merged | merge | unavailable | make sure position_ids are passed in for causal mask creation for gpt-oss | merge clean; compileall and PR-local ruff passed, but full validation unavailable due baseline ruff failures and tests_fetcher backend issue |
| huggingface#39941 | defect | merged | merge | unavailable | fixing image_utils.py todo | merge conflict was limited to imports in test_image_utils; resolved and merged validate_kwargs warning behavior; compileall and PR-local ruff passed, but full … |
| huggingface#39895 | feature | merged | merge | unavailable | Add Videoprism | merge clean; compileall and PR-local ruff passed, but full validation unavailable due pre-existing baseline ruff failures and tests_fetcher timeout/checkout is… |
| huggingface#39866 | defect | merged | merge | unavailable | make sure model.save_pretrained has the correct is_main_process | merge conflict was limited to current save_pretrained call shape; resolved by passing is_main_process plus save_safetensors; compileall and PR-local ruff passe… |
| huggingface#39794 | defect | applied | patch | unavailable | Fix ProphetNet forward to handle tuple encoder_outputs | ported ProphetNet tuple encoder_outputs conversion to current forward implementation; compileall and PR-local ruff passed, but full validation unavailable due … |
| huggingface#39793 | defect | merged | merge | unavailable | Fix DAC conversion script | merge clean; compileall and PR-local ruff passed, but full validation unavailable due baseline ruff failures and tests_fetcher checkout blockage on untracked c… |
| huggingface#39785 | defect | merged | merge | unavailable | fix mllama integration tests | merge conflict was limited to Mllama vision encoder hidden-state collection; resolved to include initial and final encoder states; compileall and PR-local ruff… |
| huggingface#39741 | defect | merged | merge | unavailable | Fix HfArgumentParser to filter out dict types from Union | merge clean; compile passed, but baseline validation already fails/unavailable due existing untracked/generated legacy files causing global ruff failures and t… |
| huggingface#39735 | defect | already_present | none | unavailable | handle multimodal models with tp_plan on the text_config | direct PR conflicted with the newer distribute_model signature, but the current branch already routes tp_plan through model.tp_plan populated from submodules/t… |
| huggingface#39698 | defect | applied | patch | unavailable | Fix exaone4 layer_types ZeroDivision/TypeError when sliding_window_pattern is None/"LLLG" | patch applied; compile and ruff on changed files passed, but full validation unavailable because baseline global ruff failures and tests_fetcher/import-structu… |
| huggingface#39697 | defect | merged | merge | unavailable | use untyped storage for dtensors due to deprecation | merge clean; validation unavailable because baseline has global ruff failures and tests_fetcher import-structure failure; compileall passed |
| huggingface#39690 | feature | applied | patch | unavailable | Allow custom hf_quantizer in from_pretrained | patch applied; compileall and narrow ruff on changed files passed, but full validation unavailable due baseline global ruff/test-fetcher failures |
| huggingface#39683 | defect | applied | patch | unavailable | Fix issue huggingface#39191 respect accelerate config to disable torch.dynamo compilation | patch applied; compileall and narrow ruff on changed file passed, but full validation unavailable due baseline global ruff/test-fetcher failures |
| huggingface#39675 | defect | already_present | none | unavailable | [BugFix]: Support dict and config file path for deepspeed | direct merge conflicted because TrainingArguments has been reorganized; current branch already types deepspeed as dict-or-path and passes it unchanged to HfTra… |
| huggingface#39674 | defect | applied | patch | unavailable | Fix loss scaling and token aggregation to use only data parallel group | patch applied; compileall and narrow ruff on changed file passed, but full validation unavailable due baseline global ruff/test-fetcher failures |
| huggingface#39625 | defect | already_present | none | unavailable | Fix: allow Union[str, dict, None] fields like deepspeed to be passed via CLI | direct merge conflicted in hf_argparser.py, but the current parser already filters dict out of Union types before argparse handling, so Union[str, dict, None] … |
| huggingface#39617 | defect | already_present | none | unavailable | Fix FSDP v1 bug: trainer incorrectly uses an unwrapped model | direct merge conflicted in Trainer setup, but the current branch already assigns the result of accelerator.prepare(self.model) to the local model and then stor… |
| huggingface#39599 | defect | applied | patch | unavailable | Fix: check TrainerState file exists before loading during resume | patch applied; compileall and narrow ruff on changed file passed, but full validation unavailable due baseline global ruff/test-fetcher failures |
| huggingface#39560 | defect | applied | patch | unavailable | fix load_model_end = true work when save_steps < eval_steps | patch applied; compileall and PR-local ruff passed, but full validation unavailable due baseline global ruff/test-fetcher failures |
| huggingface#39555 | feature | already_present | none | unavailable | [WIP] try to relax the tie_weights method | direct merge conflicted in modeling_utils.py, but current tie_weights has already been refactored to rely on expanded tied-weight mappings and no longer contai… |
| huggingface#39493 | defect | applied | patch | unavailable | [Voxtral] nit + pin correct mistral common version | ported Voxtral mistral-common dependency extras to current >=1.10.0 pin; compileall and PR-local ruff passed, global style/light validation matched failing bas… |
| huggingface#39491 | defect | applied | patch | unavailable | Fix: Skip weight initialization for quantized int8 models | skips missing-key weight initialization for quantized models; compileall and PR-local ruff passed, global style/light validation matched failing baseline |
| huggingface#39468 | defect | applied | patch | unavailable | Fix quantized model dispatch with device_map='auto' | skips accelerate dispatch_model for bitsandbytes quantized models; compileall and PR-local ruff passed, global style/light validation matched failing baseline |
| huggingface#39464 | defect | already_present | none | unavailable | Skipping initialize_weights when model is quantized | the cumulative branch already skips _initialize_missing_keys weight initialization when is_quantized is true from the prior quantized-init port |
| huggingface#39456 | defect | already_present | none | unavailable | Fix quantized model initialization for int8 dtypes | current cumulative branch already skips missing-key initialization for quantized models, covering the PR's modeling_utils fix |
| huggingface#39449 | defect | already_present | none | unavailable | Fix logger warnings in Gemma model test files | tracked Gemma/Gemma2/Gemma3/Gemma3n files no longer contain the logger.warning_once calls changed by the PR; only an existing untracked legacy Flax artifact st… |
| huggingface#39435 | defect | applied | patch | unavailable | Add a unit test for BartModel to compare eager, sdpa on one particular set of inputs | added Bart eager/SDPA mask regression test; compileall and PR-local ruff passed, targeted pytest/light validation unavailable due baseline import-structure bac… |
| huggingface#39309 | defect | merged | merge | unavailable | Fix audio pipeline with torchcodec input | merged with conflict resolution preserving current torchcodec guard and mono-channel handling; compileall passed and PR-local ruff passed, but full checker and… |
| huggingface#39264 | defect | already_present | none | unavailable | Fix: Add version check for timm to support mobilenetv5 models (fixes huggingface#39208) | current TimmWrapper already routes model creation through _create_timm_model_with_error_handling, raising an ImportError that asks users to upgrade timm for un… |
| huggingface#39257 | defect | merged | merge | unavailable | Fix to tuple conversion with config | merged with conflict resolution keeping current helper functions and applying the return_dict=True fix; compileall and PR-local ruff passed, but full checker/l… |
| huggingface#39222 | feature | already_present | none | unavailable | Enable granite 4 hybrid integration tests | current test already has GraniteMoeHybrid slow integration tests enabled against the newer ibm-granite/granite-4.0-h-tiny checkpoint with updated expected logi… |
| huggingface#39211 | defect | already_present | none | unavailable | Add mobilenet_v5 stub implementation to fix "Unknown Model" error | current cumulative branch already addresses mobilenetv5_300m_enc loading via TimmWrapper _create_timm_model_with_error_handling and timm-version upgrade guidan… |
| huggingface#39206 | defect | applied | patch | unavailable | fix: filter None router logits in Qwen3 MoE and handle empty router logits (huggingface#39203) | guarded Qwen3 MoE load-balancing loss against empty router-logit tuples; compileall and PR-local ruff passed, while full checker/light validation remain unavai… |
| huggingface#39183 | feature | applied | patch | unavailable | Add a 'chat' extra | added a chat optional extra for current rich/requests chat CLI dependencies and documented installing it; compileall and setup.py ruff passed, global validatio… |
| huggingface#39150 | feature | already_present | none | unavailable | Efficient Expert Weight Fusion for Moe deepseek v3 | current DeepseekV3MoE already routes through DeepseekV3NaiveMoe/MixtralExperts for fused/vectorized expert execution; the PR older manual weight-stacking imple… |
| huggingface#39140 | feature | already_present | none | unavailable | feat(trainer): emergency checkpointing on crashes & SIGTERM/SIGINT | current Trainer already has opt-in enable_jit_checkpoint signal-based checkpointing with CheckpointManager coverage; PR emergency checkpointing conflicts with … |
| huggingface#39108 | defect | applied | patch | unavailable | Disable static cache on certain MoE models | disabled fullgraph compilation for DeepseekV3 and Dots1 MoE classes; compileall and PR-local ruff passed, full validation remained baseline-unavailable |
| huggingface#39103 | defect | applied | patch | unavailable | Fix audio-related config naming for Gemma3n | direct merge conflicted in Gemma3n files; applied the small naming fix locally; compileall passed but full validation is unavailable because baseline ruff alre… |
| huggingface#39047 | feature | merged | merge | unavailable | RFC: refactor causal lm loss to handle lm_head in loss function | merged with a local conflict resolution in loss_utils for current typing style; compileall passed but full validation is unavailable because baseline ruff alre… |
| huggingface#39037 | defect | merged | merge | unavailable | fix kosmos2 tests | merged with local conflict resolution adapting the is_causal fix to the current Cache/past_key_values API; compileall passed but full validation is unavailable… |
| huggingface#39012 | defect | already_present | none | unavailable | [WIP] Fix DeepseekV3ModelTest::test_torch_compile_for_training | direct merge would drag a large upstream-main merge; the intended DeepSeekV3 change (iterate only matched experts, avoiding token_indices.numel dynamic-shape g… |
| huggingface#39009 | feature | merged | merge | unavailable | Add submodels support check function | merged cleanly; compileall passed but full validation is unavailable because baseline ruff already fails on pre-existing untracked/generated files |
| huggingface#38999 | defect | already_present | none | unavailable | Use deep copies instead of shallow copies for bbox_embed in GroundingDINO decoder (huggingface#37333). | current GroundingDINO already builds separate bbox_embed modules via a comprehension and manages tie keys for the shared/non-shared modes; direct PR conflict a… |
| huggingface#38908 | defect | already_present | none | unavailable | Add support to use config dtype in HybridChunkedCache | current HybridChunkedCache already uses getattr(config, 'torch_dtype', dtype); direct merge conflicts with a moved get_mask_sizes/HybridCache region from the o… |
| huggingface#38888 | defect | merged | merge | unavailable | continue to fix distributed_type from TPU to XLA in LM examples (huggingface#38652) | merge clean; validation unavailable because baseline already fails checkers on pre-existing untracked legacy/generated files and light validation tests_fetcher… |
| huggingface#38886 | feature | merged | merge | unavailable | Allow compile with bnb | merge completed with straightforward conflict resolution in bnb quantizers to keep current weight conversion methods and add PR is_compileable properties; full… |
| huggingface#38884 | defect | merged | merge | unavailable | Llama 4 conversion fix for moe models | merge completed with simple conflict resolution choosing the PR moe_args None check; validation unavailable due baseline checkers/tests_fetcher failures; narro… |

Rejected / not included records (379)

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 |

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

| huggingface#39544 | documentation | skipped | 🌐 [i18n-KO] Translated feature_extractors.md to Korea | category not configured for this cumulative branch |
| huggingface#39541 | feature | aborted | Add Muon optimizer implementation and integration | codebase moved on: Muon optimizer integration conflicts across the current optimizer registry, Trainer optimizer factory path, TrainingArguments optimizer enum, import-utils optional backend plumbing… |
| huggingface#39534 | feature | aborted | Add Beit3 model | codebase moved on: the BEiT3 new-model PR conflicts in generated auto-configuration and image-processing mappings, and porting the 2.8k-line non-modular model would require reconciling current genera… |
| huggingface#39517 | documentation | skipped | 🌐 [i18n-KO] Translated compressed_tensor.md to Korean | category not configured for this cumulative branch |
| huggingface#39480 | feature | aborted | Add model arcinstitute state | codebase moved on: new state/state-transition model conflicts in generated auto configuration/modeling registries, and the added non-modular model files use older style/import conventions; safely lan… |
| huggingface#39466 | documentation | skipped | README: Update Bert Japanese model card | category not configured for this cumulative branch |
| huggingface#39403 | feature | aborted | Add Vocos model | codebase moved on: large Vocos/Vocos-EnCodec model integration conflicts across generated auto registries, processing_utils, config-attribute checks, docs TOC, and audio utilities; safely landing the… |
| huggingface#39357 | documentation | skipped | Update docstring for glm4v | category not configured for this cumulative branch |
| huggingface#39353 | defect | aborted | fix colpali mapping | codebase moved on: PR adds ColPali to the old VLMS checkpoint-conversion path and a class mapping, but current modeling_utils removed that VLMS conversion logic in favor of newer load-state/weight-ma… |
| huggingface#39303 | documentation | skipped | Fix critical typos in code example | category not configured for this cumulative branch |
| huggingface#39297 | defect | aborted | Fix bug with deepspeed and accelerator args in training_args.py | codebase moved on: PR narrows deepspeed and accelerator_config annotations to Optional[str] on an older TrainingArguments layout, while current code intentionally supports dict | str | None with _V… |
| huggingface#39293 | feature | aborted | Add T5LA models | codebase moved on: new T5LA model integration conflicts across generated auto mappings, docs TOC, model package exports, checker rules, and a deleted fx utility; landing it would require regenerating… |
| huggingface#39251 | defect | aborted | Fix slow test_moshika_greedy_unconditional_fp16 | codebase moved on: PR depends on an older cache_utils/generation/Moshi implementation, while the cumulative branch has a newer cache layer architecture and Moshi prepare_inputs_for_generation overrid… |
| huggingface#39236 | feature | aborted | added moment_p sampling | codebase moved on: moment-p sampling touches generation exports, configuration validation, logits processors, docs, and tests that now conflict with current generation APIs/registries; integrating th… |
| huggingface#39212 | documentation | skipped | Add Ukrainian translation of README.md | category not configured for this cumulative branch |
| huggingface#39209 | other | skipped | Standardize FSMT class naming: PretrainedFSMTModel → PreTrainedFSMTModel | category not configured for this cumulative branch |
| huggingface#39109 | defect | aborted | Fix: rename 'eval_strategy' to 'evaluation_strategy' in TrainingArgum… | codebase moved on: TrainingArguments in current v5 intentionally uses eval_strategy throughout docs, dataclass fields, post-init validation, and helper methods; the PR renames it back to evaluation_s… |
| huggingface#39084 | feature | aborted | Refactor gemma3n | codebase moved on: current Gemma3n is generated from strict modular/config code and already has conflicting audio naming updates; the PR refactors configuration, audio encoder placement, residual fie… |
| huggingface#39046 | defect | aborted | Fix deprecated max_size parameter handling in DETR image processors | codebase moved on: DETR-family image processors in current tree have been rewritten around ImageProcessorKwargs/SizeDict helpers, while the PR edits older processor internals and even adds backup/ad-… |
| huggingface#38991 | feature | aborted | Remove return_dict kwarg from all the models | 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; direct merge conflicts immediat… |
| huggingface#38988 | defect | aborted | Check docstring inside modular files as well | codebase moved on: PR combines checker changes with widespread modular/model docstring rewrites from an older generated state; current tree has deleted/moved fast image processor and args_doc artifac… |
| huggingface#38962 | other | skipped | Update test_candidate_generator.py | category not configured for this cumulative branch |
| huggingface#38959 | documentation | skipped | Updated the model card for wav2vec2-phoneme | category not configured for this cumulative branch |
| huggingface#38958 | documentation | skipped | Updated model card for wav2vec2-conformer | category not configured for this cumulative branch |
| huggingface#38957 | documentation | skipped | Update wav2vec2-bert model card | category not configured for this cumulative branch |
| huggingface#38956 | documentation | skipped | Updating model card for wav2vec2 | category not configured for this cumulative branch |
| huggingface#38955 | documentation | skipped | docs: Musicgen melody model card | category not configured for this cumulative branch |
| huggingface#38926 | documentation | skipped | Clarify Python and framework version support in installation.md | category not configured for this cumulative branch |
| huggingface#38923 | feature | aborted | Remove deprecated max_size support from YOLOS image processor | codebase moved on: current YOLOS image processor is generated from modular_yolos and the inherited DETR v5 image processor kwargs path; the PR edits older generated init/preprocess/from_dict code… |
| huggingface#38893 | defect | aborted | Fix/deprecate max size conditional detr | codebase moved on: Conditional DETR image processors have been split/refactored into torch and PIL v5 processors generated from current image-processing APIs; the PR edits older max_size/from_dict/pr… |
| huggingface#38877 | documentation | skipped | DOC: Clarify attention_mask usage in BertModel forward method | category not configured for this cumulative branch |
| huggingface#38861 | feature | aborted | Add SamImageProcessorFast with 4x performance improvement | codebase moved on: current v5 SAM image processing uses the new TorchvisionBackend/PilBackend split and auto image processor mappings are generated/dict-based; the PR adds an older BaseImageProcessor… |
| huggingface#38859 | feature | aborted | Add MobileViT fast image processor | codebase moved on: current MobileViT image processing is already rewritten for the v5 TorchvisionBackend/PilBackend split and dynamic import structure; the PR adds an older BaseImageProcessorFast imp… |
| huggingface#38839 | other | skipped | [DO NOT MERGE] Testing saftensors 0.6.0 | category not configured for this cumulative branch |
| huggingface#38810 | feature | aborted | Add kwargs support in WhisperForConditionalGeneration | codebase moved on: Whisper forward signatures now use v5 type annotations, Cache, can_return_tuple, and kwargs forwarding; the PR is based on the older Optional/return_dict/head_mask API and its loss… |
| huggingface#38805 | feature | aborted | Add Dust3R | codebase moved on: adding Dust3R conflicts with current auto-generation/modular model structure and v5 image processor backend; the PR modifies fully generated auto mapping files, adds an older fast … |
| huggingface#38793 | documentation | skipped | Fix Typos in Comments and Improve Clarity | category not configured for this cumulative branch |
| huggingface#38786 | documentation | skipped | Provide clearer instructions on how to specify target language. | category not configured for this cumulative branch |