bf16 halves the matting model's activation memory -- its full-resolution
attention is the pipeline's memory peak -- with visually identical alpha
(measured: 0 px alpha deviation > 0.25 on samples; cross-check veto
behaviour unchanged, region IoU 0.984).
Guarded by a hardware check so it never lands on a slow emulation path:
CUDA requires torch.cuda.is_bf16_supported(); CPU requires the same
oneDNN native-bf16 gate PyTorch uses for matmul routing (AVX512-BF16/
AMX). Without support it warns and falls back to fp32 -- on a Zen2 EPYC
the fallback kernels measured 17-370x slower than fp32, so silent bf16
there would be a performance landmine.
Segmentation stays fp32: torchvision deform_conv2d (used by BiRefNet)
has no bf16 CPU kernel, and the segmenter is not the memory peak.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
lhk229
dfc85566c7