huggingface/transformers
Readiness B🤗 Transformers: the model-definition framework for state-of-the-art machine learning models in text, vision, audio, and multimodal models, for both inference and training.
Python · ★ 162,546 · 100 open issues synced
Contributor readiness
Contributor readiness: 75/100 (grade B)
- Active development
- README
- Contributing guide
- Code of conduct
- CI on pull requests
- Contribution manifest
- Responsive maintainers
Maintainer signals
Maintainers respond in ~<1 day (median first response, measured from real issue sampling; 65% of sampled issues got a response).
This project has no valid contribution manifest yet, so maintainer preferences are unknown.
Local setup
- CI: add-model-like.yml, anti-slop.yml, assign-reviewers.yml, benchmark.yml, benchmark_v2.yml, benchmark_v2_a10_caller.yml, benchmark_v2_mi300_caller.yml, build-ci-docker-images.yml, build-docker-images.yml, build-nightly-ci-docker-images.yml, build-past-ci-docker-images.yml, build_documentation.yml, build_pr_documentation.yml, check-workflow-permissions.yml, check_failed_tests.yml, check_tiny_models.yml, codeql.yml, collated-reports.yml, dependents-smoke-test.yml, doctest_job.yml, doctests.yml, extras-smoke-test.yml, get-pr-info.yml, get-pr-number.yml, model_jobs.yml, model_jobs_intel_gaudi.yml, new_model_pr_merged_notification.yml, nightly-integration-failure-triage-caller.yml, pr-ci-caller.yml, pr-ci-post-dashboard-link.yml, pr-ci-security-gate.yml, pr-repo-consistency-bot.yml, pr_build_doc_with_comment.yml, pr_slow_ci_suggestion.yml, push-important-models.yml, release-conda.yml, release.yml, self-comment-ci.yml, self-nightly-caller.yml, self-nightly-past-ci-caller.yml, self-past-caller.yml, self-scheduled-amd-caller.yml, self-scheduled-amd-mi250-caller.yml, self-scheduled-amd-mi300-caller.yml, self-scheduled-amd-mi355-caller.yml, self-scheduled-caller.yml, self-scheduled-flash-attn-caller.yml, self-scheduled-intel-gaudi.yml, self-scheduled-intel-gaudi3-caller.yml, self-scheduled.yml, slack-report.yml, ssh-runner.yml, stale.yml, trl-ci-bot.yml, trufflehog.yml, update_metdata.yml, upload_pr_documentation.yml
Links
Open issues
Sign in to see which issues fit your skills- #47276 Misleading synced_gpus warning message in continuous batching
intermediate · unassigned · bug
- #47274 stable diffusion in tensorflow
intermediate · unassigned
- #47273 Cohere ASR chunk reassembly can crash on empty chunks and accept inconsistent metadata
intermediate · unassigned
- #47272 Avoid repeated Cohere ASR encoder projection during cached generation
intermediate · unassigned
- #32902 StoppingCriteria for Repetition
intermediate · unassigned · Feature request
- #45488 LlamaTokenizer in v5 overrides tokenizer.json's ByteLevel pre-tokenizer with Metaspace, silently breaks DeepSeek V3/R1 family
intermediate · unassigned
- #46516 [Continuous batching] Tracker for issues and feature requests
intermediate · assigned · Performance · Inference
- #46570 AutoModel loading fails with ROCm when model has custom CUDA kernels
intermediate · unassigned
- #45175 Feature request: Add EfficientViT-SAM (efficientvitsam) to Transformers
intermediate · unassigned · Feature request
- #47217 VL models fail when their image placeholder is in user text.
intermediate · unassigned · bug
- #35532 RagTokenizer Missing patch_token_id, patch_token, and encode Functionality
intermediate · unassigned · Good Second Issue · Feature request
- #42179 Add TileLang Kernel Support
intermediate · unassigned · Feature request
- #47265 [serge] integration failure triage - 2026-07-11
intermediate · assigned
- #46222 คิดดี ทำดี เจอแต่สิ่งดี สาธุ 365.ai.th
intermediate · unassigned · WIP
- #46868 Sequence Parallelism Disabled for Evaluation
intermediate · unassigned
- #46552 Wav2Vec2CTCTokenizer in v5 treats the word delimiter as a special token, which leaks into get_special_tokens_mask and convert_ids_to_tokens
intermediate · unassigned
- #47254 `CheckpointError` with PEFT + DeepSpeed ZeRO-3 + gradient checkpointing
intermediate · unassigned · bug
- #47246 Incorrect inter-chunk recurrence in `NemotronHMamba2Mixer.torch_forward` (slow path). the Mamba2 fix from #35154 was never propagated to `nemotron_h`
intermediate · unassigned · bug
- #47177 Tar-slip / zip-slip arbitrary file write → RCE ...
intermediate · unassigned · bug
- #47228 Sam3 bf16 compilation regression (v5.8.0 -> v5.9.0)
intermediate · unassigned · bug
- #39215 _load_rng_state after get_batch_samples may break training reproducibility when dataloader has random operations
intermediate · unassigned · bug
- #44273 Lazy loading is not working properly
intermediate · assigned · bug
- #47225 orthogonal_ weight init crashes with `geqrf_cuda not implemented for BFloat16` under v5 bf16 on-device init (RWKV and any QR-based init)
intermediate · unassigned
- #47085 Fast (Rust-backed) tokenizers aren't thread-safe anymore
intermediate · unassigned · bug
- #47205 Standardize test image fixture loading for COCO images
intermediate · unassigned
- #46506 Different results for PPDocLayoutV3 on CPU and CUDA
intermediate · unassigned · bug
- #46512 from_config(dtype=...) builds weights in the requested dtype but leaves model.config.dtype stale
intermediate · unassigned
- #46577 Add RTDetrForSegmentation
intermediate · unassigned · Feature request
- #46491 CodeLlama tokenizer strips one leading space on encode→decode round-trip (regression vs v4)
intermediate · unassigned · bug
- #47145 Silent corruption: FP8 compressed-tensors "frozen" checkpoints load without applying weight_scale
intermediate · unassigned · bug
Showing 30 of 100 open issues — the rest are on GitHub.