MPS (Apple Silicon): .float() on device tensors crashes with float64 error — use .to(torch.float32)

[musicalplatypus]

Environment

• Platform: Apple Silicon (MPS)
• PyTorch MPS backend
• tinyml-tinyverse r1.4

Bug

tinyml_tinyverse/common/utils/utils.py calls .float() on tensors after .to(device) in multiple training functions. On Apple Silicon (MPS), .float() resolves to float64, which MPS does not support:

Affected lines

All occurrences of the pattern .to(device[, non_blocking=True]).float() in utils.py, e.g.:

# Line ~1452 (and 15 other sites)
data = data_feat_ext.to(device).float()  # crashes on MPS


Fix
Replace .float() with .to(torch.float32) — explicit dtype is the PyTorch-recommended pattern for cross-device code:

data = data_feat_ext.to(device).to(torch.float32)  # works on MPS, CUDA, CPU
Workaround
Applied locally to the installed package (16 sites in utils.py).