One landing instruction: the adapter tensor tensor_name (a manifest key),
which half it is, the device destination address (a pool-slot sub-region
base, already pool + slot*slot_bytes + a_off|b_off), and the geometry the
convert/repack needs. rank is the adapter’s real r; max_rank the pool’s
padded rank.
Which half of a LoRA pair a target lands. A is copied contiguous into the
head of the padded [max_rank, in] region; B is row-repacked from stride
r to stride max_rank into [out, max_rank].
Host F16/F32/BF16 → BF16, byte-for-byte matching the disk adapter loader
(load_adapter_safetensors): half::bf16::from_f32 (round-to-nearest-even)
for the float conversions. Any other dtype is a hard error — a PEFT adapter
is only ever F32 (default), F16, or BF16.
The final host bytes to copy_h2d for a target, given the raw on-wire
tensor bytes (as landed in the bounce). A: convert only. B: convert + repack.
Factored out (un-gated) so the byte-identity logic is unit-testable off the
RDMA path — the verbs loop calls this so tested and shipped logic agree.
Row-repack a BF16 B tensor [out_dim, r] into the pool’s padded
[out_dim, max_rank] layout: per row, r BF16 elements copied from stride
r to stride max_rank; pad columns stay zero. Byte-identical to the disk
pack’s B repack. Returns out_dim * max_rank * 2 bytes.