Cumulative defect fixes from recent Transformers PRs

[evalstate]

Cumulative defect fixes from recent Transformers PRs

This PR is generated by the all-defects mergeability flow. It accumulates defect-fix PRs from huggingface/transformers that could be applied cleanly to the current base.

• Source branch: all-defects
• Base: evalstate/transformers:main
• Head: f342edbf6c
• PRs classified: 134
• PRs with terminal state: 134
• Applied/merged/already-present defect fixes: 65
• Aborted defect fixes: 5
• Validation failures reverted: 1
• Non-defect skipped: 63

Status counts

• aborted: 5
• already_present: 6
• applied: 1
• merged: 58
• skipped: 63
• validation_failed: 1

Category counts

• defect: 71
• documentation: 12
• feature: 30
• other: 21

Validation

Each applied defect fix was followed by the configured lightweight validation profile:

1. compileall -q src/transformers
2. utils/checkers.py ruff_check,ruff_format,init_isort,sort_auto_mappings
3. utils/tests_fetcher.py ... && pytest ... when impacted pytest targets are selected

Note: this is intentionally not an end-to-end or slow-test validation pass.

Details

A detailed status table is posted as a PR comment and is also available locally in:

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

[evalstate]

All-defects flow status

Processed terminal records: 134

PR	Category	Status	Method	Validation	Summary
huggingface#45679	other	skipped			TST Run fast PEFT tests in normal CI — not a defect fix
huggingface#45678	defect	merged		passed	Fix shared config mutation issue in flash_attn_from_config — merged cleanly; validation commands exited 0
huggingface#45677	defect	merged		passed	No serving in quality docker image — merged cleanly; validation commands exited 0
huggingface#45675	defect	merged		passed	Fix UnboundLocalError in shard_and_distribute_module for replicated parameters — merged cleanly; validation commands exited 0
huggingface#45673	feature	skipped			Laguna XS.2 implementation — not a defect fix
huggingface#45671	defect	merged		passed	Update latest revision for Phi-4-multimodal test — merged cleanly; validation commands exited 0
huggingface#45670	defect	merged		passed	[nit] glmasr should be in AutoModelForMultimodalLM — merged cleanly; validation commands exited 0
huggingface#45669	documentation	skipped			zero_shot_object_detection ValueError fix for python 3.13 — not a defect fix
huggingface#45668	feature	skipped			[GGUF] Add support for Qwen3.5 MoE (qwen35moe arch) — not a defect fix
huggingface#45667	other	skipped			chore(typing): add ty type checking for 3 pipeline files — not a defect fix
huggingface#45681	defect	merged		passed	Restore TokenizersBackend override for DeepSeek V3/R1 tokenizer dispatch — merged cleanly; validation commands exited 0
huggingface#45680	defect	aborted			change got reverted — overlaps with already merged huggingface#45681 DeepSeek tokenizer dispatch fix; conflicts in same AutoTokenizer logic and regression tests, so applying both would be duplicate/speculative
huggingface#45666	feature	skipped			Extended n-to-1 kernel fusion via KernelConfig — not a defect fix
huggingface#45665	defect	merged		passed	Fix pageable H2D copies in Gated DeltaNet PyTorch fallback — merged cleanly; validation commands exited 0
huggingface#45664	documentation	skipped			Doc translate to Persian(farsi) — not a defect fix
huggingface#45662	defect	merged		passed	Fix EP + FSDP2: experts silently overwritten by rank-0 broadcast — merged cleanly; validation commands exited 0
huggingface#45661	defect	merged		passed	[Weight Converter] More fine-grained mappings on classes, scoping for every transforms (including weight converter) — merged cleanly; validation commands exited 0
huggingface#45660	documentation	skipped			[docs] cpu offloading — not a defect fix
huggingface#45659	documentation	skipped			[docs] dtype — not a defect fix
huggingface#45658	defect	merged		passed	Fix NameError: PeftConfigLike triggered by PreTrainedModel.__init_subclass__ — merged cleanly; validation commands exited 0
huggingface#45655	defect	merged		passed	Fix the order of cls.config resolution — merged cleanly; validation commands exited 0
huggingface#45654	feature	skipped			[CB] Refactor any model-related code in a separate class — not a defect fix
huggingface#45653	feature	skipped			[CB] Better overall script and decode bucketting — not a defect fix
huggingface#45652	defect	merged		passed	Fix colmodernvbert tests — merged cleanly; validation commands exited 0
huggingface#45651	defect	merged		passed	[Trainer] Optimize LengthGroupedSampler computation with select_columns and tqdm — merged cleanly; validation commands exited 0
huggingface#45650	defect	merged		passed	Fix KeyError for flash_attn in import_utils.py on Python 3.13Fix — merged cleanly; validation commands exited 0
huggingface#45649	defect	merged		passed	Fix OOM regression for FSDP2 + cpu_ram_efficient_loading on large models — merged cleanly; validation commands exited 0
huggingface#45648	defect	merged		passed	Fix SDPA inference tolerances for MPS backend — merged cleanly; validation commands exited 0
huggingface#45645	defect	merged		passed	Fix xdist collisions for captured_info artifacts and preserve CI debug logs — merged cleanly; validation commands exited 0
huggingface#45643	feature	skipped			Add DeepSeek V4 — not a defect fix
huggingface#45642	defect	merged		passed	Fix trust_remote_code local cache collisions for local models (huggingface#45632) — fixed local trust_remote_code cache key collisions and validated cleanly
huggingface#45641	defect	merged		passed	Fix NameError in serving CLI due to conditional import asymmetry — serving CLI conditional import NameError fix merged and validated cleanly
huggingface#45640	feature	skipped			🚨🚨🚨 [Trainer] Default to FSDP2, simplify API around fsdp + fsdp_config — not a defect fix
huggingface#45639	defect	aborted			Make patched testing debug logs xdist-safe — codebase moved on: xdist captured_info handling was already merged via huggingface#45645 with different helper names and notification artifact aggregation; this PR conflicts in testing_u
huggingface#45638	feature	skipped			Add Multi-Token Prediction (MTP) support for Qwen3.5 — not a defect fix
huggingface#45637	feature	skipped			Add Multi-Token Prediction (MTP) support for Qwen3.5 — not a defect fix
huggingface#45635	other	skipped			qa: speed up dtype regex weight load + reduce dtype tests to 3 random — not a defect fix
huggingface#45634	feature	skipped			DeepGEMM BF16, isolation, refactor — not a defect fix
huggingface#45633	other	skipped			CircleCI with torch 2.11 — not a defect fix
huggingface#45631	other	skipped			chore: bump doc-builder SHA for main doc build workflow — not a defect fix
huggingface#45630	feature	skipped			Add new model: Kimi2-6 — not a defect fix
huggingface#45629	other	skipped			Allow more artifacts to be download in CI — not a defect fix
huggingface#45628	defect	merged		passed	[MistralCommonBackend] Soften validation mode and apply_chat_template arguments check — Mistral common backend validation softening merged cleanly; validation commands exited 0
huggingface#45627	defect	merged		passed	Processing Utils: honor pre-built sub-processor kwargs in from_pretrained — AutoProcessor prebuilt sub-processor kwarg fix merged cleanly; validation commands exited 0
huggingface#45626	feature	skipped			[Model] Add PP-FormulaNet Model Support — not a defect fix
huggingface#45625	defect	merged		passed	Add supports_gradient_checkpointing to NemotronHPreTrainedModel — NemotronH gradient checkpointing support flag merged cleanly; validation commands exited 0
huggingface#45624	defect	merged		passed	Skip failing offloading tests — Gemma4 offloading test skip merged cleanly; validation commands exited 0
huggingface#45623	other	skipped			Glm5 change — not a defect fix
huggingface#45622	defect	merged		passed	Fix peft constructors — PEFT constructor fix merged cleanly; validation commands exited 0
huggingface#45621	feature	skipped			Better Grouped GEMM + EP — not a defect fix
huggingface#45620	defect	aborted			Fix TypeError in video_processor_class_from_name when torchvision is not installed — codebase moved on: PR branch appears based on an obsolete fork and replays broad repository changes while only intending a video proces
huggingface#45619	defect	merged		passed	Remove unnecessary generate warnings — generation warning cleanup was already present in cumulative branch; validation commands exited 0
huggingface#45618	feature	skipped			Add MTP speculative decoding via MTPCandidateGenerator — not a defect fix
huggingface#45617	feature	skipped			Add Multi-Token Prediction (MTP) inference support — not a defect fix
huggingface#45616	feature	skipped			Add DeepSeek V4 — not a defect fix
huggingface#45615	defect	merged		passed	fix(qianfan_ocr): add XPU expectations — Qianfan OCR XPU test expectation fix merged cleanly; validation commands exited 0
huggingface#45614	defect	validation_failed		failed	Add missing requests dependency to transformers[serving] — ruff_format failed after merge: setup.py would be reformatted
huggingface#45613	feature	skipped			[New Model] Add MiniCPM3 support — not a defect fix
huggingface#45612	documentation	skipped			[docs] update model cards — not a defect fix
huggingface#45611	defect	merged		passed	Raise clear error for problem_type="single_label_classification" with num_labels=1 — invalid single-label classification config error merged cleanly; validation commands exited 0
huggingface#45610	defect	aborted			Fix configuration reading and error handling for kernels — codebase moved on: conversion_mapping registry has since gained Laguna and generated mapping changes around qwen3_5_moe_text/pp_chart2table; PR intent is small F
huggingface#45609	feature	skipped			make it possible to ser/deser HF MoE models with torchao — not a defect fix
huggingface#45608	documentation	skipped			Python code in model docs — not a defect fix
huggingface#45607	other	skipped			Add regression test for Gemma4 audio relative positional range — not a defect fix
huggingface#45604	feature	skipped			Agent first cli with skill — not a defect fix
huggingface#45606	defect	merged		passed	[gemma4] infer from config instead of hardcoding — clean merge; validation commands passed
huggingface#45605	defect	merged		passed	Processing Utils: continue when content is a string — clean merge; validation commands passed
huggingface#45603	defect	merged		passed	[Auto] Pass kwargs to fixed_cross_entropy (cluster-43240-3): merged 2 of 2 PRs — clean merge of clustered fixed_cross_entropy kwargs fix; validation commands passed
huggingface#45602	defect	merged		passed	[AMD CI] Fix expectations for Gemma3n — clean merge; validation commands passed
huggingface#45601	defect	merged		passed	fix: compute auxiliary losses when denoising is disabled in D-FINE — clean merge; validation commands passed
huggingface#45599	defect	merged		passed	qa: more lazy loading — clean merge; validation commands passed
huggingface#45682	defect	merged		passed	FIX Restore LoRA hotswapping functionality — LoRA hotswapping fix merged cleanly; validation commands exited 0
huggingface#45681	defect	merged		passed	Restore TokenizersBackend override for DeepSeek V3/R1 tokenizer dispatch — TokenizersBackend dispatch fix was already present in cumulative branch; validation commands exited 0
huggingface#45680	defect	aborted			change got reverted — codebase moved on: tokenizer regression is already covered by the newer 45681 change in the cumulative branch; this older alternative conflicts in AutoTokenizer dispatch and overlapping auto tokeniz
huggingface#45679	other	skipped			TST Run fast PEFT tests in normal CI — not a defect fix
huggingface#45678	defect	merged		passed	Fix shared config mutation issue in flash_attn_from_config — shared config mutation fix was already present in cumulative branch; validation commands exited 0
huggingface#45677	other	skipped			No serving in quality docker image — not a defect fix
huggingface#45675	defect	merged		passed	Fix UnboundLocalError in shard_and_distribute_module for replicated parameters — tensor-parallel UnboundLocalError fix was already present in cumulative branch; validation commands exited 0
huggingface#45673	feature	skipped			Laguna XS.2 implementation — not a defect fix
huggingface#45671	defect	merged		passed	Update latest revision for Phi-4-multimodal test — Phi-4 multimodal test revision fix was already present in cumulative branch; validation commands exited 0
huggingface#45670	defect	merged		passed	[nit] glmasr should be in AutoModelForMultimodalLM — auto-model mapping fix was already present in cumulative branch; validation commands exited 0
huggingface#45598	defect	merged	merge	passed	Align latest model attention function dispatch — attention dispatch alignment merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45597	feature	skipped			Add Granite 4.1 Vision (granite4_vision) — not a defect fix
huggingface#45596	defect	merged	merge	passed	fix 2 failed test cases for blt model on XPU — BLT XPU test expectation fix merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45595	feature	skipped			Add unified Cache-layer management for GLM-5 DSA Indexer keys — not a defect fix
huggingface#45594	defect	merged	merge	passed	fix(utils): Resolve backbone utils test regressions — backbone utility test regression fix merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45592	defect	merged	merge	passed	fix padding side issue for fast_vlm tests — fast_vlm padding-side test fix merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45591	defect	merged	merge	passed	[nemotron_h] respect _no_reinit flag on dt_bias and out_proj.weight — Nemotron-H _no_reinit guard fix merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45590	defect	already_present	none	passed	fix huggingface#45588: guard s_aux against None in flash_attention_forward — s_aux None guard was already present in cumulative branch with equivalent formatting; attempted merge conflicted only on that already-applied line and was
huggingface#45589	defect	merged	merge	passed	Fix AttributeError on s_aux=None in flash_attention_forward — flash_attention s_aux None guard merged cleanly as an empty tree merge because equivalent fix was already present; validation commands exited 0
huggingface#45587	documentation	skipped			[docs] cb memory management — not a defect fix
huggingface#45586	feature	skipped			Add Audio-Visual Flamingo model — not a defect fix
huggingface#45585	other	skipped			qa: bumped mlinter and allow local override — not a defect fix
huggingface#45583	other	skipped			Update dev version — not a defect fix
huggingface#45582	defect	merged	merge	passed	generate: drop stale num_return_sequences warning on continuous batching path — stale continuous-batching num_return_sequences warning fix merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45581	other	skipped			Add automated reviewer assignment script — not a defect fix
huggingface#45580	feature	skipped			[Privacy Filter] Add model — not a defect fix
huggingface#45579	other	skipped			Update assign_reviewers.py — not a defect fix
huggingface#45581	other	skipped			Add automated reviewer assignment script — not a defect fix
huggingface#45580	feature	skipped			[Privacy Filter] Add model — not a defect fix
huggingface#45579	other	skipped			Update assign_reviewers.py — not a defect fix
huggingface#45578	defect	merged	merge	passed	Remove attribute_map from GptOssConfig — GptOssConfig attribute_map removal merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45577	defect	merged	merge	passed	Allow for registered experts from kernels hub — registered expert kernel loading fix merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45576	documentation	skipped			docs(pipeline): fix num_workers docstring default from 8 to 0 — not a defect fix
huggingface#45575	defect	already_present	none	passed	fix(generation): remove stale warning for num_return_sequences in paged generate — equivalent stale num_return_sequences warning removal is already present from the earlier cumulative generation fix
huggingface#45574	feature	skipped			Fix typos — not a defect fix
huggingface#45573	defect	merged	merge	passed	fix transformers + torchao nvfp4 serialization — merged torchao NVFP4 serialization fix cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45572	other	skipped			refactor(Dots1): drop Dots1MoE override to pass (inherits from DSV3 MoE) — not a defect fix
huggingface#45570	defect	merged	merge	passed	Fix whisper long-form generation when eos_token_id is a list — merged Whisper eos_token_id list handling fix cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45569	feature	skipped			Proper nemotron H and 3 and 2 — not a defect fix
huggingface#45568	defect	merged	merge	passed	Gemma4: fix failed test cases — merged Gemma4 failing-test fixes cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45567	other	skipped			Move some conversion mappings to PrefixChange — not a defect fix
huggingface#45566	defect	merged	merge	passed	fix: raise clear error when tokenizer config uses v5 list format on older versions — merged tokenizer v5 extra_special_tokens list guard cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45565	defect	already_present	none	passed	fix: remove stale num_return_sequences warning in paged generate — equivalent stale num_return_sequences warning removal is already present from the earlier cumulative generation fix
huggingface#45564	defect	merged	merge	passed	Gemma3n and Gemma4 cannot use rotary kernel — Gemma3n/Gemma4 rotary kernel disablement merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45562	defect	merged	merge	passed	Updated the image cache for Paddle models according to the latest API — Paddle model cached image/API test updates merged cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45560	documentation	skipped			Update torchao usage for XPU and CPU — not a defect fix
huggingface#45559	defect	already_present	none	passed	Drop noisy generate warnings when do_sample=False (or num_beams=1) — generation config user-set-attribute warning suppression is already present in a refined form; direct PR head also includes many unrelated branch diffe
huggingface#45558	feature	skipped			feat(trainer): log individual losses from loss_dict — not a defect fix
huggingface#45556	documentation	skipped			Add image processors refactor to v5 migration guide — not a defect fix
huggingface#45555	other	skipped			perf: avoid recomputing rotary_emb for each layer in some Google and ModernBERT models — not a defect fix
huggingface#45554	documentation	skipped			[docs] multi-turn tool calling — not a defect fix
huggingface#45553	documentation	skipped			[docs] per-request sampling params — not a defect fix
huggingface#45552	defect	applied	patch	passed	Remove warnings for modernbert — applied minimal auto_docstring modular file inclusion to suppress ModernBERT import warnings; validation commands exited 0 with no pytest targets selected
huggingface#45551	feature	skipped			Add ForSequenceClassification heads for the OLMo family — not a defect fix
huggingface#45550	other	skipped			Add runner selection for mi325 GPU type — not a defect fix
huggingface#45549	defect	merged	merge	passed	fix: apply channel averaging correctly in audio feature extractors — merged audio channel-averaging axis fixes cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45548	defect	merged	merge	passed	Fix EP + DeepSpeed ZeRO-3 loading via accelerate launch — merged EP plus DeepSpeed ZeRO-3 loading guard cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45547	defect	already_present	none	passed	Add disable_mmap kwarg to from_pretrained with hf-mount auto-detection — disable_mmap and hf-mount safetensors fallback are already present in the cumulative branch; direct merge only conflicted on the same load_state_di
huggingface#45546	feature	skipped			feat: Add GGUF loading support for Llama 4 (text) — not a defect fix
huggingface#45544	defect	already_present	none	passed	fix table update versions — Python 3.10 patch-version table-update condition is already present; direct merge conflicted only with newer transformers-mlinter dependency version 0.1.1
huggingface#45543	other	skipped			ci: OTEL support — not a defect fix
huggingface#45541	defect	merged	merge	passed	Fix local_files_only tokenizer fallback when tokenizer files are missing (Issue 45538) — merged local_files_only tokenizer fallback fix cleanly; validation commands exited 0 with no pytest targets selected
huggingface#45540	defect	merged	merge	passed	Fix cross-attention cache layer type for T5Gemma2 long inputs — merged T5Gemma2 cross-attention cache fix cleanly; tree already had equivalent contents and validation commands exited 0 with no pytest targets selected