[arav-agarwal2]

Companion to mlcommons/endpoints_policies#160, which moves disaggregated,
expert_parallel, tensor_parallel, pipeline_parallel, data_parallel and
config_summary out of the top level of system_desc.json and into each
node_types entry.

_point_parallelism read only the top level, with the structured
config_summary object as a fallback. It now reads the node-type placement
first — that being the current format — and falls back to the old one, so a
bundle in either shape works and one part-way through the move resolves field by
field. Within a node type, its own config_summary is the fallback, exactly as
the SUT-wide one was.

Two consequences that needed handling, not just the lookup

1. The cross-point identity check would have broken.
_system_desc_identity strips the fields that legitimately vary between points
of a frontier — and it strips them by top-level key:

return {k: v for k, v in data.items() if k not in _PER_POINT_SYSTEM_DESC_FIELDS}

With those fields nested, that no longer reaches them. Any frontier that tunes
its replica size per point — which §4.5.3's own C.3 worked example does — would
then differ on node_types and fail system-description-consistency with
differing: node_types. They are now stripped at both depths, so a bundle in
either placement compares on the same identity. Without this the format move
would have failed real submissions, which is why it's in this PR rather than a
follow-up.

2. Node types that disagree yield None rather than a pick. §4.5.3 computes
one A_replica = TP·PP·EP for the whole SUT, so a Parallelism cannot carry two
values for a field. Normalising against a replica size no node type actually has
would be worse than skipping nodes-used — which is already what happens for a
bundle that declares no parallelism at all. Making §4.5.3's arithmetic per node
type is the real fix and is its own piece of work; point_power.py already
documents the gap it leaves ("for a heterogeneous system the split across sets is
not knowable").

Also

• NodeType declares the six fields. ConfigSummary and _ConfigSummaryStr
  move above it so a node type's config_summary takes the same type as the
  SUT-wide one — a reorder, no change to either definition.
• disaggregated is true if any node type is disaggregated.

Verification

1805 tests pass, 15 new. The new ones cover: the node-type placement read;
top-level still read; config_summary still the last fallback; node-types
winning over the top level; the two placements mixing per field; a node type's
own field beating its own config_summary; agreeing node types resolving;
disagreeing ones declining; any disaggregated node type disaggregating the SUT;
declaring nothing still None; and four on identity — nested parallelism and
nested config_summary excluded from identity, real hardware differences still
counted, and a malformed node_types passed through rather than raising.

🤖 Generated with Claude Code

[github-actions]

MLCommons CLA bot All contributors have signed the MLCommons CLA ✍️ ✅