Datasets:

---

nvidia

/

aisimulate-fpm-dataset

Like 0

Follow

NVIDIA 68.8k

Summary

Integrate the four AgentX FPM captures into the dataset's standard configuration
layout, following the review request. This replaces the temporary
proposals/agentx-fpm-20260911/ storage with discoverable measurement-only leaves.
The 32 old proposal files are relocated/replaced in this PR; the previous commit
retains their original form. No main-branch data is deleted or overwritten.

Layout and catalogs

• Standard data/<org>--<model>/<system>/<framework>/<framework-version>/<parallelism>/
  leaves, flat configuration manifests, measurement manifests and catalog entries.
• Complete list[list[FPM]] files are truth directly under measurements/.
  Incomplete groups, indexes, collection evidence and alignment/processing
  provenance are under measurements/provenance/agentx-job-<job>/.
• The data are unchanged from sequence-aligned PR revision
  0d1bc03e053e5e327d2893943c5d2e99a993abed: 708,535 complete iterations,
  1,873,391 complete-group rank observations, and 23,759 incomplete observations.
  All 1,897,150 retained FPM records remain available.
• Four new configurations bring the active catalog from 16 to 20 configurations.
  The existing 343 data files, 38 FPM libraries and eight historical snapshots
  remain byte-identical.
• SGLang package versions were not captured. Their explicit version identity is
  git-71de97b264b0-fpm-ed18d64951b9, with full engine/patch/image provenance;
  no release version or AISim runtime commit is fabricated.

Additive contracts and tooling

• Configuration manifests permit fpm: [] when validated measurement truth
  exists. Empty and summary-only external configurations remain invalid.
• forward-pass-measurement-v1 supports pre_grouped_rank_lists with file-level
  producer, expected ranks, authority and hashed alignment provenance.
• scripts/pre_grouped.py streams the lists and normalizes stored membership to
  canonical synchronizedIteration. It does not rematch ranks or impose an
  8 ms receive window. Derived alignment remains explicitly labeled as such.
• The validator checks complete rank membership, positive finite timing,
  duplicate observations, counts, hashes and declared selection windows.
• Catalog construction preserves fields absent from the first row, avoiding
  Arrow silently dropping new representation/grouping metadata. Validation also
  checks manifest/JSON/Parquet metadata agreement.
• scripts/ingest_agentx.py is a checksum-pinned, collision-checked migration.
  Base Playground rebuilds fail closed instead of erasing external measurements.
• Collection/processing documentation is in docs/agentx-fpm-20260911.md;
  original preparation code is retained under scripts/agentx/.

Preserved collection semantics

The four runs are native FPM-on captures with pinned model/image/FPM revisions,
AgentX replay settings and SD/G2 configuration. Native vLLM uses PR52061 emission,
not Dynamo's InstrumentedScheduler. Formal receipt windows remain exactly 3600
seconds, with GLM's PDT export correctly converted to UTC. Warmup, cooldown/drain
and zero-time heartbeats are excluded; positive-time empty-rank work is retained.

SGLang alignment uses audited full-capture ordinals. vLLM alignment uses monotone
sequence matching with independent reference and sensitivity checks. There is no
fixed 8 ms membership cutoff, no fabricated missing rank and no duration-based
record deletion. vLLM post-alignment duration similarity is not independent proof
of timing accuracy. This migration changes storage/metadata only, not grouping or
measurement values.

The repository remains private. No credentials, prompts, generated responses,
model weights, full logs or raw request traces are added. Historical proposal
paths survive only in immutable-source provenance.

Validation

• python -m unittest discover -s scripts -p 'test_*.py': 16 passed.
• python -m unittest discover -s scripts/agentx -p 'test_prepare.py': 12 passed.
• ruff check on new ingestion/normalization/test modules: passed.
• All six JSON Schemas checked; all 28 configuration and measurement-manifest
  pairs validate; 20 actual canonical normalized sample records validate.
• Full dataset validator: passed on both the unchanged baseline and the
  integrated dataset. It streamed all 708,535 new complete iterations.
• All 343 existing data files and every moved FPM/index payload were checked
  for byte equality. Original full-source validation covered all 1,897,150
  retained records; the source validation report is preserved in the ingestion audit.
• Documentation relative links resolve. Integration accounting and validation
  are recorded in audit/agentx-ingestion.json,
  audit/agentx-integration-validation.json and audit/validation-report.txt.

This PR supplies dataset-side schema, discovery, normalization and validation.
It does not claim an external FPM Gym deployment has already adopted the new
representation; no external Gym deployment was modified or tested.

Signed-off-by: hongkuanz hongkuanz@nvidia.com

Updated per request: removed the fixed 8 ms membership rule. SGLang/B300 now has 98,227 complete groups (incomplete: 20,483 -> 0); native vLLM/B200 has 68,181 complete groups (incomplete: 19,193 -> 3,654). All 1,897,150 retained FPM payloads are unchanged; DP1 data files are byte-identical. Alignment uses audited full-capture SGLang ordinals and independently cross-checked monotone vLLM sequences. Twenty-six ambiguous vLLM observations remain explicit singletons; no missing rank is fabricated. Full original-payload validation, monotonicity/uniqueness checks, 12 tests and checksums pass. README, indexes, manifests and alignment reports have been regenerated. IDs remain inferred, and vLLM duration agreement after matching is not independent evidence of timing accuracy.

Updated per the layout/schema review: all four captures now live under standard model/system/framework/version/parallelism leaves with flat configuration manifests, measurement manifests and catalog entries. Removed the active proposals prefix. Added additive measurement-only and pre-grouped-list schema/validator support; complete truth is flat under measurements, incomplete groups and collection/alignment evidence are under provenance. Grouping and FPM/index payload bytes are unchanged. Full dataset validation passes on both old and integrated data; 28 tests pass; 343 existing data files and all historical snapshots are unchanged. See docs/agentx-fpm-20260911.md and audit/agentx-integration-validation.json. Dataset-side integration is complete in this PR; external Gym deployment support is not claimed.