Issue evidence · Run · huggingface/diffusers

GGUF: enable_sequential_cpu_offload() raises KeyError: None

GitHub issue: https://github.com/huggingface/diffusers/issues/14691

Reproduced

GGUFParameter looks up its quant type unguarded, so re-wrapping a parameter during accelerate's offload raises KeyError: None and GGUF models larger than VRAM cannot run.

What Badgr ran

Ran the issue's minimal reproduction on CPU with diffusers 0.41.0: build GGUFParameter(torch.zeros(32,32), quant_type=8), then GGUFParameter(p, requires_grad=False).

What came back

  • The first call worked and kept quant_type 8.
  • The re-wrap raised KeyError(None), as the issue shows.

This is the failing re-wrap itself, not the full FLUX pipeline with sequential offload. The reporter's fix is to forward quant_type when parameters are re-wrapped.

Checked 2026-10-15 in Badgr’s local development environment.

← All issue evidence