Build the unfused projections on the meta device
unfuse_projections reconstructs to_q/to_k/to_v by pointing them at views into
the fused weight, but built the layers on the default device first -- so every
reconstructed projection was materialised and Kaiming-initialised in float32
only to have its parameters immediately replaced. On a 5.5B-parameter bf16 Flux
transformer that is 3.4 GiB of allocation and 4.0s of wasted init; on `meta` it
is nothing measurable and 18ms.
The Wan-family transformers inherit this method, so they get the same saving.
The accompanying test checks that no parameter is left on meta, which is the
new hazard: one that is not replaced by a view would break the module.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
