Read system_power.json as Appendix E defines it

[arav-agarwal2]

Summary

Reads system_power.json using the schema in Appendix E, added by policies PR #126. §9.1's Power descriptor row now checks against Appendix E and its validation rules (E.7).

#126 is still open, and Appendix E itself flags two schema choices as WG open items: sourcing each value separately, and the node_sets array. This PR follows the current text of #126.

PR #126 also changes the power model, not just the file shape, so reading new files the old way would give the wrong total:

System Power = Major + Other + Published_node_power + Scale_out_switch_power
Major        = Σ component_sum sets: nodes × (CPU + Accelerator + Scale-up) + NICs where counted
Other        = overhead_fraction × Major          (0.30 liquid, 0.50 air)
Published    = Σ published_system: nodes × node figure  +  Σ node_scaling: P_rack × Y/N

What changed

• Schema: one entry per set of identical nodes (node_sets), each costed by component_sum, published_system or node_scaling. Every power figure is a sourced value (value_w/value_kw/value_pj, a source_type and a source). A figure the submitter asserts without a public source is not a valid source_type and fails to load.
• Arithmetic (E.5):
  • Scale-out NICs are major components and take the overhead fraction.
  • Published node power and rack switch power are wall figures, so no overhead is applied to them.
  • A declared_provisioned_power replaces the computed total. Only its own sourcing decides the estimated tag; the component block under it is just a cross-check.
• Appendix D defaults (power_defaults.py) fill values left out of the file, wherever a default can be chosen mechanically:
  • D.3: accelerator TDP by model.
  • D.2: CPU TDP by architecture and core count. Cores come from §8.2's node_types, or from the model name.
  • D.4: NIC power, and reference switches by cabling.
  • Any filled value sets MLC Estimated Power (a power-estimated warning).
  • Scale-up has no fallback, because D.1's two references depend on the link protocol, which the file doesn't declare.
• E.7 rejections, each reported as its own power-descriptor error:
  • a missing required field
  • a bad source
  • cooling disagreeing with system_desc.json
  • node_scaling where N ≤ Y
  • switches that don't carry the required bandwidth
  • cabling missing when there is a scale-out fabric
  • a computed block that disagrees with the checker's own numbers
• §4.5.2's NIC rules are enforced alongside E.7:
  • A multi-node formula system must count its NICs.
  • A published node figure must not count them again, unless excluded_from_published_power gives evidence that the figure leaves them out.
• A descriptor that is rejected produces no provisioned_power_kw, so nothing is normalized against it.
• The fixtures and their generator now write Appendix E files, and the generator is still idempotent.

Stacked on #99

This branch is based on #99, which fixes the Llama name that #94's merge broke on main. Without #99, 10 tests fail on main. Merge #99 first; GitHub then retargets this PR to main.

Worth a look

• aggregate_bandwidth_tbps units. The name says Tb/s, but C.8, C.9 and E.6.2 all treat it as TB/s (14.4 × 8 × 5 pJ = 576 W). scale_out.required_bandwidth_tbps is genuinely Tb/s. The code follows the worked examples; the spec should settle one unit.
• Old flat descriptors no longer validate. Appendix E replaces the §4.5.2 field-group reading. There's no compatibility path.
• Appendix D lives in the release, for now. D says the values are revisited at least quarterly; data/seed_sets.yaml is the pattern for moving the table out of the release when that starts.

Test plan

• The arithmetic reproduces every total the spec gives: E.6.1 300.30 kW, E.6.2 14.50 (component sum 15,114 W), E.6.3 146.80 (149.16 with active optical), E.6.4 111.00, and C.11's formula path 161.94.
• Unit tests for every E.7 rejection, the NIC rules, defaults and gaps, disagreement with the computed block, and rounding.
• End-to-end tests on valid_standardized: cooling disagreement, a D.2 default read from the system description, and an unknown node ensemble.
• pytest: 1092 passed. ruff check and ruff format are clean. regenerate_fixtures.py --check is a no-op.

🤖 Generated with Claude Code

[github-actions]

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

[anandhu-eng]

We may need to check if its single or multi node system that's being submitted. For multi node systems, the rule mandates. ref: https://github.com/mlcommons/endpoints_policies/blob/nvashutoshd_power_example/endpoints_rules.md#e4-scale-out-fabric