Datasets:

---

nvidia

/

aisimulate-fpm-dataset

Like 2

Follow

NVIDIA 69.3k

Define one canonical synchronized forward-pass iteration with nested rank measurements. Both physical source forms normalize to this shape; their grouping provenance differs.

Canonical shape

Required fields are version, source_kind, iteration_id, producer, grouping, expected_dp_ranks, complete, max_rank_wall_time, and rank_measurements. collector is optional globally and required for listener receive-time-window grouping.

• rank_measurements is the authoritative per-rank evidence, ordered by ascending dp_rank.
• max_rank_wall_time is the cached default scoring latency. Consumers may derive another policy from rank_measurements and record that choice in their result provenance.
• A complete iteration contains exactly one measurement for every rank in expected_dp_ranks and requires max_rank_wall_time == max(rank_measurements[*].wall_time).
• An incomplete iteration retains available ranks with max_rank_wall_time: null.

Source and grouping forms

• self_benchmark uses grouping.method: benchmark_id. The producer supplies the shared benchmark_id, point, expected ranks, and one FPM per rank. The physical source iteration_groups[].wall_time normalizes to canonical max_rank_wall_time after validation.
• A one-rank event stream uses grouping.method: single_rank.
• A multi-rank listener stream uses grouping.method: listener_receive_time_window.

Listener timestamps are assigned independently for each received event, so exact _recv_ts equality is not a join rule. The canonical grouping instead records:

• engine role and raw _recv_ts unit;
• the clock-correction method, reference clock, rank-to-clock groups, and signed offsets;
• the corrected window start, size, half-open boundary, and distinct-dp_rank rule; and
• the component that performed the derived grouping.

Every contributing rank retains its raw _recv_ts. File order and rank-local counter_id are not cross-rank identities. The GLM DEP16 campaign is represented by four corrected node clocks followed by an 8 ms window of distinct ranks.

Physical representations

• Rank observation streams remain one rank record per JSONL line.
• Self-benchmark files retain their producer iteration_groups representation.
• Importers may materialize canonical synchronized iterations as one JSONL record per iteration.

Standalone rank records are directly evaluable only when attention_dp_size == 1; multi-rank truth requires a complete synchronized iteration.

Validation passes for JSON Schema Draft 2020-12 meta-validation, canonical self-benchmark and listener-window cases, physical self-benchmark groups, rejection of the former exact-timestamp method, and the full current dataset (15 configurations, 22 measurement bundles, and 163 source measurement files).