Closed RaulPPelaez closed 4 months ago
Tensornet enables tf32 by default. We find these affect the accuracy too greatly in some situations so I am disabling them out of caution (which is the default in pytorch). User code can request TF32 usage by writing this after importing torch:
torch.set_float32_matmul_precision("high") torch.backends.cuda.matmul.allow_tf32 = True
Tensornet enables tf32 by default. We find these affect the accuracy too greatly in some situations so I am disabling them out of caution (which is the default in pytorch). User code can request TF32 usage by writing this after importing torch: