A staged model’s manifest: the geometry the client needs to reconstruct a
byte-identical WeightStore. Published as length-prefixed JSON right after
the client’s model request. The per-shard (base, rkey) MR handles ride the
verbs handshake separately (see the module doc).
One tensor’s placement inside a staged model, mirroring the safetensors
header exactly: offset_in_shard is the ABSOLUTE file offset (8-byte size
prefix + header + the tensor’s data-section start), len is the raw
contiguous byte count (data_offsets[1] - data_offsets[0]). The client
RDMA-READs exactly [shard_base + offset_in_shard .. + len).
Rail selection for a tensor under dual-rail striping: tensor N is served
over rail N % n_rails. Factored out (un-gated) so the striping is unit-
testable off the RDMA path — the client’s read loop calls this so the tested
logic and the shipped logic are the same. n_rails is clamped to >= 1.
Absolute peer virtual address of a tensor’s first byte: the shard’s whole-
file REMOTE_READ MR base plus the tensor’s ABSOLUTE in-shard offset (the
safetensors data-section offset, which already includes the 8-byte size
prefix + header). The client RDMA-READs [addr .. addr + len). Factored out
(un-gated) so the address math is unit-testable off the RDMA path.