Expand description
transpose_head_to_token op-diff harness — see crate::op_diff.
Reorders [tokens, heads, dim] head-major data to token-major layout.
Pure data movement (exact), so accelerator NMSE should be ~0.
transpose_head_to_token op-diff harness — see crate::op_diff.
Reorders [tokens, heads, dim] head-major data to token-major layout.
Pure data movement (exact), so accelerator NMSE should be ~0.