Repository Issue Activity (beta)

#14114 `_flash_3_varlen_hub` backend cannot handle non-contiguous mask

#14122 `from_single_file()` support missing/incomplete for Ideogram4 and Krea2 transformers

#14104 Flux-family attention: flash-attn and other backends fail in an autocast context

#14124 [Version 0.39.0]AttributeError: 'CLIPTextModel' object has no attribute 'text_model' WHEN SET CKIP_SKIP NOT NONE

#13793 Support for Stable Audio 3 Medium

#14090 Fix model offloading and training tests + prevent examples timeout

#14117 `DiffusionPipeline.download()` breaks with huggingface_hub>=1.22.0 in offline mode (IncompleteSnapshotError)

#14013 A bug on the set_timesteps function of the FlowMatchEulerDiscreteScheduler

#14025 MagCache on Wan 2.2 Dual-Transformer Pipelines: Incorrect Step Accounting and Limited Effectiveness on a 4-Step Distilled Model (e.g. Wan2.2)

#14067 uv pip install -e ".[test]" fails on Python 3.12 due to incompatible numba/llvmlite resolution

#13999 Flash Attention 3 varlen wrapper assumes output is a tuple

#14052 Fix mutable default argument in lora_base.py

#14097 Regression (#13485) Broken TorchAO Compat

#14002 Support Original Checkpoint-Compatible PEFT Adapters

#14103 Custom AttentionProcessor lost/fallback occurring after forward pass

#14063 Kandinsky5 pipeline does not support device_map=balanced

#14095 SuperCompress: Free open-source LLM prompt compression - cut token costs by ~65%

#14094 Reduce prompt costs with SuperCompress compression

#14037 `HookRegistry._child_registries_cache` goes stale after `enable_cache`/`disable_cache` → `ValueError: No context is set`

#14088 Handling of .alpha in from-comfy converters

#14072 [Bug] Layerwise casting tests fail on MPS backend due to float8 unsupported dtype

#14004 PixArtSigmaPipeline: no LoRA loading support

#14086 Add SeFi-Image pipeline

#14085 Backport workflow-hardening fix (excessive-permissions) to 3 release branches

#14056 Modular pipeline: Krea 2