Issues

Issues

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

  huggingface/diffusers#15264 · 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 · 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 · 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 · opened on Nov 12, 2026

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

  huggingface/diffusers#15252 · 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 · 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 · opened on Nov 12, 2026

• Automatically reusing components

  huggingface/diffusers#15248 · 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 · opened on Nov 11, 2026

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

  huggingface/diffusers#15242 · opened on Nov 11, 2026

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

  huggingface/diffusers#15239 · opened on Nov 11, 2026

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

  huggingface/diffusers#15233 · opened on Nov 10, 2026

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

  huggingface/diffusers#15227 · opened on Nov 9, 2026
  3

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

  huggingface/diffusers#15224 · opened on Nov 9, 2026

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

  huggingface/diffusers#15221 · opened on Nov 9, 2026

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

  huggingface/diffusers#15220 · opened on Nov 9, 2026

• _tp_plan / _cp_plan: allow entries for modules that are not always present

  huggingface/diffusers#15218 · opened on Nov 9, 2026 · Release 0.44.0
  8

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

  huggingface/diffusers#15211 · opened on Nov 8, 2026
  3

• set_ip_adapter_scale(float) raises "Cannot assign 1 scale_configs to 2 IP-Adapter" when several IP-Adapters are loaded

  huggingface/diffusers#15210 · opened on Nov 8, 2026
  1

• [modular] Per-workflow input signature in the docstring of blocks with _workflow_map

  huggingface/diffusers#15208 · opened on Nov 8, 2026
  3

• UniPCMultistepScheduler.set_timesteps(sigmas=...) raises TypeError for a list when use_flow_sigmas=True

  huggingface/diffusers#15207 · opened on Nov 7, 2026
  1

• device_map="balanced" drops all but the last CPU-overflow component, which then loads without offload hooks and breaks inference

  huggingface/diffusers#15204 · opened on Nov 7, 2026
  3

• FLEX attention silently ignores is_causal=True when no mask is supplied

  huggingface/diffusers#15195 · opened on Nov 6, 2026
  2

• HookRegistry.remove_hook drops the next hook's new_forward (breaks FirstBlockCache with group offloading)

  huggingface/diffusers#15193 · opened on Nov 6, 2026

• Per-token FlowMatch Euler steps use the opposite velocity sign

  huggingface/diffusers#15185 · opened on Nov 5, 2026
  2