Issue evidence · Run · huggingface/diffusers

Group offloading with use_stream=True fails on torchao int8 version=2

GitHub issue: https://github.com/huggingface/diffusers/issues/14433

Reproduced

apply_group_offloading with use_stream=True fails with "cannot pin 'torch.cuda.CharTensor'" for torchao int8 weights.

What Badgr ran

Ran the issue's self-contained reproducer (a 4-block stack of torchao Int8WeightOnlyConfig(version=2) linears under apply_group_offloading) on an RTX 4090 through Badgr with torch 2.9.0+cu126, torchao 0.17.0 and diffusers 0.41.0.

What came back

  • use_stream=True, low_cpu_mem_usage=True: RuntimeError: cannot pin 'torch.cuda.CharTensor' only dense CPU tensors can be pinned.
  • use_stream=False, low_cpu_mem_usage=True: OK.
  • use_stream=True, low_cpu_mem_usage=False: OK.

Matches the issue's three outputs. This is the small synthetic stack from the issue, not MiniMax-H3 itself; the cu126 build of torch 2.9.0 was used because the cu128 build failed on that host's driver. The two OK configurations are the workaround.

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

← All issue evidence