Datasets:

---

nvidia

/

aisimulate-fpm-dataset

Like 0

Follow

NVIDIA 68.7k

Summary

Add a separate Kimi K3 configuration on 8 GB300 GPUs with TP8 + DCP8, attention DP1, MoE TP8/EP1, FP8 KV cache, and the original 32-request engine limit. Preserve the existing Kimi TEP8 configuration and every existing data artifact.

• Add 669 primary balanced FPM points: 338 prefill and 331 decode with real attention KV. Decode KDA/conv state is random; prefill follows the native state path.
• Keep 14 synthetic-attention-KV decode boundary points in a comparator and 56 nonuniform prefill points, including exact request layouts, in calibration provenance. The two native source projections contain all 739 measured points; the subsets do not add observations.
• Record collection method dynamo_native_self_benchmark, source kind self_benchmark, scheduler wall-time units, per-row repeat policy, source hashes, engine settings, and both producer revisions. Calibration data is not published as independent evaluation truth.

DCP is a TP-overlapping dimension, not independent CP: this is 8 GPUs, not 64. Extend aic_fpm_forward_perf v6 with optional dcp while keeping every existing v6 artifact unchanged. This configuration explicitly records dcp=8 in its manifest, canonical sidecar selector, and Parquet identity. Missing DCP remains unrecorded, not implicitly DCP1. GPU count remains in manifest/provenance. The dataset is compatible with companion Gym MR !39 (b7ffc4a2d3dc7432c83fb097d1e3cf686ce32c52); AISimulate DCP prediction support remains a separate requirement.

New provenance is sanitized: no machine paths, internal hostnames/IPs, service endpoints, private registry URL, raw engine logs, model weights, or source dataset text are included. Native result values and synchronized iteration groups are preserved exactly. Original and sanitized source hashes are separately recorded and pinned in the reviewed import policy. The validator binds actual published JSON bytes to fixed sanitized SHA-256 values even when source, output, and local provenance hashes are rewritten together. Unresolved runtime metadata is left unknown rather than inferred.

Validation

• Both source jobs reject coordinated source/iteration/Parquet tampering after all dependent hashes are regenerated. Sidecar tests reject missing/wrong DCP and manifest-only fields.

• Companion Gym MR !39: 300 tests passed, one real-AISim dependency skip. A local reader/discovery smoke check against the actual prepared leaf accepts v6, preserves DCP8/CP1 and 8 physical GPUs, shows both calibration FPM variants with no independent truth, and reports native DCP predictions unavailable.

• Verified the original local package checksums and native completeness/coverage for both successful runs.

• Exact native-source-to-Parquet comparison for every published balanced point; nonuniform source rows retained; finite positive timings, request counts, token totals, repeat policy, DP grouping, KV regimes, and topology checked.

• Validated all configuration/measurement JSON schemas and the catalog schema.

• Full dataset validator passes, including all existing campaigns and catalog/provenance hashes.

• 43 dataset tests pass: DCP identity and synthetic/calibration boundary regressions plus existing pre-grouped and AgentX preparation tests. Used an isolated compatible NumPy/Numba environment.

• Rebuilt from the original local package to check reproducibility; the updated validator also accepts the unchanged baseline dataset.

• Existing data artifacts remain byte-identical. A scan of the new data/provenance found no machine paths, intranet hosts, IPs, service endpoints, or credential markers.

See docs/kimi-k3-self-benchmark-20260916.md for counts, timing semantics, compatibility limits, and validation commands.

Please extend the existing aic_fpm_forward_perf v6 schema with DCP instead of introducing the separate aisim_fpm_forward_dcp schema.

• Keep dcp separate from independent cp, and carry it through the configuration manifest, sidecar selector, and Parquet identity. DCP must divide TP and must not multiply the physical GPU count.
• Preserve existing v6 artifacts; an absent DCP field means it was not recorded. Consumers must not silently use DCP8 measurements for ordinary TP8.
• We will open a companion Gym MR to preserve and validate DCP, keep calibration-only configurations visible, and report unsupported prediction paths explicitly. AISimulate prediction support remains a separate requirement.

At Gym e822172, this PR currently raises DataError: ... declares an unsupported schema during dashboard discovery, before reaching the no-measurements case. Please coordinate the dataset change with that consumer update.

Companion Gym MR: !39 — preserve DCP identity in v6 FPM inputs. It expects the existing canonical v6 sidecar selector plus dcp; physical_gpu_count stays in manifest/provenance. Local validation: 300 tests passed, 1 real-AISim dependency skip.

Please also pin the reviewed sanitized source hashes in the import policy and add a regression test.

validate_sources() checks the claimed original_sha256 against SOURCES, but checks the actual published JSON against a hash supplied by the editable import.json. It does not bind those bytes to the reviewed original.

Reproduction in a temporary copy: change one source latency from 66.17 ms to 132.34 ms, update its synchronized group, matching Parquet, and dependent hashes, while leaving the claimed original hash and producer commit unchanged. validate_snapshot() still passes.

Add the expected sanitized SHA-256 values to the reviewed allowlist and reject changed source bytes even when the local provenance and output hashes are regenerated. This is a validation gap; I found no actual mismatch between the submitted Parquets and published source records.