Issues

• LTX-2.3 in float16: text-connector RMS norm overflows, pipeline silently returns an all-NaN video

  huggingface/diffusers#15281 · Filter by author AuthRan opened on Nov 13, 2026

• Helios transformer builds a per-token (batch, seq_len, 6, dim) float32 modulation tensor in every block

  huggingface/diffusers#15279 · Filter by author chsu-amd opened on Nov 13, 2026

• add_noise / get_velocity in DDPM-family schedulers add no noise at the first timesteps for fp16/bf16 samples

  huggingface/diffusers#15277 · Filter by author Robertboy18 opened on Nov 13, 2026

• ERNIE-Image transformer expands the adaLN modulation to the full sequence and upcasts the activations to float32 in every block

  huggingface/diffusers#15272 · Filter by author chsu-amd opened on Nov 13, 2026

• HunyuanVideo-1.5 builds a dense (seq_len, seq_len) attention mask in every transformer block

  huggingface/diffusers#15271 · Filter by author chsu-amd opened on Nov 13, 2026

• [ROCm] native attention backend: SDPA on strided K/V is several times slower on gfx1100 for long sequences

  huggingface/diffusers#15270 · Filter by author Shan2L opened on Nov 13, 2026 · Release 0.44.0
  1

• [tests] A tester mixin for running modular blocks on their own

  idea-contribution-welcome
  huggingface/diffusers#15267 · Filter by author yiyixuxu opened on Nov 12, 2026
  1

• EDMEulerScheduler.set_timesteps fails with a list of sigmas, and custom sigmas are passed through the karras ramp

  huggingface/diffusers#15266 · Filter by author MaiAditya opened on Nov 12, 2026

• Option to page-lock offloaded CPU weights in place for streamed group offloading

  huggingface/diffusers#15264 · Filter by author Shan2L opened on Nov 12, 2026

• UniPCMultistepScheduler returns NaN on the last step with final_sigmas_type='zero' (solver_type='bh1' or lower_order_final=False)

  huggingface/diffusers#15263 · Filter by author davrens101 opened on Nov 12, 2026

• Shap-E renderer fine pass re-runs the NeRF MLP on the 64 coarse samples per ray

  huggingface/diffusers#15260 · Filter by author chsu-amd opened on Nov 12, 2026

• Streamed group offloading with low_cpu_mem_usage=True re-pins every group on every forward pass

  huggingface/diffusers#15256 · Filter by author Shan2L opened on Nov 12, 2026

• LuminaFeedForward upcasts to float32 around SiLU, which F.silu already does on CUDA

  huggingface/diffusers#15252 · Filter by author chsu-amd opened on Nov 12, 2026

• Lumina-Next repeats avoidable work on every denoising step (expanded attention mask, 384 x 384 RoPE table on CPU)

  huggingface/diffusers#15251 · Filter by author chsu-amd opened on Nov 12, 2026
  1

• enable_group_offload(offload_type="block_level") fails on LTX / LTX-2 / LTX-2.5 pipelines (vae.decode, audio_vae.decode, LTX2VocoderWithBWE, and nested block_modules)

  huggingface/diffusers#15250 · Filter by author Shan2L opened on Nov 12, 2026

• Automatically reusing components

  idea-contribution-welcome
  huggingface/diffusers#15248 · Filter by author yiyixuxu opened on Nov 11, 2026

• [MiniMax-H3, Pika H3 Flex] ref2va fails on a reference soundtrack that is not at 32 kHz without torchaudio

  huggingface/diffusers#15245 · Filter by author christopher5106 opened on Nov 11, 2026

• LLaDA2: EOS early stopping runs every remaining block to the step limit in batched generation

  bug
  huggingface/diffusers#15242 · Filter by author yyu-260 opened on Nov 11, 2026

• Stable Audio 3 attention repeats RoPE cos / sin and q / k norm work on every call

  huggingface/diffusers#15239 · Filter by author chsu-amd opened on Nov 11, 2026

• Docs: 17 dead external links (verified ×3) — upstream content removed, needs maintainer decisions

  huggingface/diffusers#15233 · Filter by author Hisernberg opened on Nov 10, 2026

• How should ModularPipeline handle a component that failed to load?

  idea-contribution-welcome
  huggingface/diffusers#15227 · Filter by author yiyixuxu opened on Nov 9, 2026 · Release 0.44.0
  4

• Wan transformer block makes float32 copies of hidden_states in its residual adds

  huggingface/diffusers#15224 · Filter by author klin2024 opened on Nov 9, 2026

• FP32LayerNorm makes two full-size float32 copies on every call

  huggingface/diffusers#15221 · Filter by author klin2024 opened on Nov 9, 2026

• RoPE helpers make float32 copies of q / k on every call

  huggingface/diffusers#15220 · Filter by author klin2024 opened on Nov 9, 2026

• unload_textual_inversion(token) leaves the extra tokens of a multi-vector embedding behind, so it can't be loaded again

  huggingface/diffusers#15211 · Filter by author sohailshk opened on Nov 8, 2026
  3