viable/strict/1780750020: [MPS] fix attention compilation on nightly (#186399)
🔒
https://github.com
«Fix attention compilation on nightly:
import torch
from torch.nn.functional import scaled_dot_product_attention as sdpa
q, k, v = (torch.randn(1, 1179, 16, 128, device="mps", dtype=torch.bfloat16).transpose(1, 2) for _ ...»
Automatische Weiterleitung...
1.5s