Devices and precision¶
pipeline("t2i", backend="sd-v1-5", model="checkpoints/sd-v1-5", device="cpu") # default
pipeline("t2i", backend="sd-v1-5", model="checkpoints/sd-v1-5", device="mps") # Apple silicon
pipeline("t2i", backend="sd-v1-5", model="checkpoints/sd-v1-5", device="cuda") # NVIDIA
Asking for a device that is not available raises an error that lists what is.
Precision¶
dtype accepts torch.float32 (default), torch.float16 or torch.bfloat16 where a model
supports them:
import torch
pipeline("t2i", backend="sd-v1-5", model="checkpoints/sd-v1-5", device="cuda", dtype=torch.float16)
What is claimed where¶
| Device | Precision | Claim |
|---|---|---|
| CPU | float32 | Parity: bit-identical to the reference for the recorded checkpoints |
| MPS | float32 | Smoke test: finite output, right size, difference from CPU recorded |
| CUDA | float32 | Not tested yet |
| any | float16 / bfloat16 | No parity claim |
Fast is not the same as identical
Different hardware sums floating-point numbers in a different order, so GPU output can differ from CPU output by a few pixel values. imagehub records that difference where it was measured and does not call it parity.
The full numbers are in the validation record.