Expand description
K3 Megatron TP plan. supports_tp is true — the umbrella loader used to
refuse --tp-size 2. Full slice_for_rank bind lives with the weight
loader; this slice publishes the plan + the fail-fast flag.
Head counts on config are already per-rank. Full sizes = local * tp.
Functions§
- supports_
tp - K3 does tensor-parallel. Callers must not fail-fast
supports_tpfalse. - tensor_
plan (kind, full_out, full_in)for one checkpoint key.