[Afloat16]

Problem

For a quantized Int8Params, to() creates a new parameter and preserves its packed data, but only copies SCB when the caller explicitly supplies a device. Valid no-device calls such as weight.to(), weight.to(dtype=torch.int8), and weight.to(non_blocking=True) therefore discard the row scales.

Replacing an embedding or a linear layer's weight with that result breaks its first quantized forward pass.

Change

Carry the existing SCB into the new parameter before handling an explicit device move. Continue binding CB to the returned parameter's data, and forward non_blocking when moving the scales.

Validation

• 15 new cases cover no-argument, dtype-only, non-blocking-only, and explicit-device conversions, including meta-device scale preservation.
• Native Embedding8bit and Linear8bitLt cases replace the weight before the first forward, so cached matmul statistics cannot hide lost metadata. Their outputs, downstream gradients, and two SGD updates match the unchanged models exactly.
• Original source: 12 failures and 3 passing controls.
• Patched source: all 15 new cases pass; 62 existing embedding/Int8Params cases pass and one existing unsupported CPU case is skipped.
• Source-built AVX2 CPU kernels, PyTorch 2.14.0, Ruff check and formatting.

Exact production commit: eeaa2f804eda7d04d4326097e6140429b01e247a.
Validation: https://github.com/Afloat16/bitsandbytes/actions/runs/43401310219