Rebuild from AISimPlayground 731947a: DSV4-Pro dep8 r6pin and GLM-5.2 GB200 DEP16 with canonical truth
Rebuilds the dataset from AISimPlayground 731947aabdcb5ae8658fd16744022ca94e407cd3 on top of main 32725a9 (protocol/provenance refactor, Qwen removal, runtime dtype annotations, unified forward-pass iteration measurements #7). Two FPM additions, one regeneration with scripts/manage_dataset.py import --replace (pyarrow 25.0.0, the version that wrote the current parquets, so no untouched parquet changed bytes).
1. DeepSeek-V4-Pro GB300 dep8 r6pin (campaign dsv4pro-gb300-fpm, Kimi Zhao). The prefill lanes were re-measured with worker pinning (fpm/dep8-kvon-full4096-j3521395-3545915, fpm/dep8-kvoff-full4096-j3521396-3545916); the decode rows are byte-identical to the r4 libraries, the 4,398 prefill coordinates are unchanged and their latencies re-measured (median 310.6 ms to 285.2 ms). New current snapshot aisim-commit-unknown-6cdbf517 for data/deepseek-ai--DeepSeek-V4-Pro/gb300/vllm/0.26.0/dep8: kv-on primary, kv-off comparator, protocol forward-pass-record-v1 with 7 truth-only derivatives from predictions/dep8-r6pin/ flat under measurements/ (six decode tiers c1 to c64 plus prefill, kv-on scores, same selection as tep8 r9), the r6pin windows.tsv and serving configs under measurements/provenance/, and the two pinned lanes' resolved configs under fpm/provenance/ linked from the kv-on and kv-off metadata. The runtime-dtype-annotation-v1 correction applies to the new libraries like the rest of the campaign. The previous dep8 snapshot moves to data/history/.../dep8/aisim-commit-unknown-d7ffaefd/ unchanged (same snapshot id and file set; its kv-on library is now role historical, matching tep8 r3).
2. GLM-5.2-NVFP4 GB200 DEP16 (new campaign glm52-gb200-dep16-disagg-fpm-fidelity, Yiming Liu; AISimPlayground MR !17, !20, !21, !22). New leaf data/nvidia--GLM-5.2-NVFP4/gb200/vllm/0.28.0/dep16 (tp 1, dp 16, moe_ep 16, nvfp4 weights, fp8 KV, model revision aec724e8, snapshot aisim-commit-82ffa025): one primary library of 8,229 rows (fpm/dep16-kvon-pairs-realseed-dg-chunks-j7024594-7025934-7052366/): a decode cell on the full paired batch axis (68 rows, every CUDA-graph capture size and its +1 neighbour, 1,025 points, 987 real KV and 38 fake-fallback) and a real-seeded prefill cell on every 8,192-token chunk boundary (7,204 points). Its dtype columns already carry the runtime labels, so no correction policy is declared. The earlier 12-row decode pair is discarded with a reason. The measurement bundle is a forward-pass-measurement-v1 bundle whose truth is reconstructed at import (see below): measurements/fpm_iterations_prefill.jsonl.gz (67,220 iterations, 182 complete, from 96,819 rank records) and measurements/fpm_iterations_decode.jsonl.gz (280,194 iterations, 91,163 complete, from 3,056,367 rank records). The raw rank-event streams (one per engine role, 16 DP ranks), their per-rank provenance record, and the bucketed tables, reports, record samples and aiperf exports are supporting evidence under measurements/provenance/disagg16-truth-v2-j6924441/. The raw streams are retained on the maintainer's request and are what the validator cross-checks the canonical record count against; the ten per-node and per-tier sample files are now strict subsets of those streams and can be dropped if the "redundant copies of rank streams" rule is read strictly.
Canonical iteration reconstruction (new importer step, LISTENER_WINDOW_NORMALIZATIONS). For campaigns declared in that table, the importer decompresses each rank-event stream, estimates the clock offset of every clock group (one per node, 4 ranks) against the reference clock from long busy steps shared in lockstep (median of matched receive-time deltas; method shared-long-step-offset-median-v3, a line-by-line port of the campaign's truth_table.py; on the decode engine no busy step exceeds 150 ms, so the offsets are declared zero rather than estimated, and every decode record says so), sorts records by corrected receive time, forms half-open 8 ms windows anchored at the earliest ungrouped record with at most one record per dp_rank, and emits one synchronizedIteration per window: source_kind: rank_event_stream, producer, collector, grouping.method: listener_receive_time_window with the clock groups and offsets, window rule and window_start, expected_dp_ranks 0..15, complete, max_rank_wall_time (null when incomplete), and rank_measurements sorted by dp_rank with every raw record byte-identical to the source line (_recv_ts retained). Equivalence to the campaign's own tool was verified on both streams: identical offsets and match counts on the prefill engine (node1 +0.25 ms, node2 +0.73 ms, node3 -0.43 ms), identical ranks-present histograms (prefill 67,220 windows / 182 complete; decode 280,194 / 91,163), and on the decode engine both find no busy step above 150 ms and assume zero offsets, which every record states (status: insufficient_matches_assumed_zero). Every source record lands in exactly one iteration. The validator now parses each canonical file: version, source kind, grouping method and authority, engine role, expected ranks, sorted distinct ranks, complete consistency, max_rank_wall_time equal to the maximum rank wall time, retained _recv_ts, clock groups partitioning the ranks with a zero reference offset, and every member inside its own window; it also requires the raw stream to be retained as supporting evidence and the rank-record total to equal the raw stream's record count.
Truth aggregation wording. The campaign's scorer takes the median wall over busy ranks as a step's measured time and the maximum over busy ranks on the prediction side; on the published complete decode steps the maximum exceeds the median by about 1% at the median (p90 1.7%, p99 8%; prefill p50 0.03%). The dataset default max_rank_wall_time is therefore close but not identical to the campaign's measured step time; the campaign READMEs now say so (AISimPlayground MR !22), and the dataset README and audit notes carry the same statement.
Other importer changes. DSV4PRO_DERIVED_TRUTH_SOURCES = 25 and DSV4PRO_FPM_PROVENANCE_COUNTS = (15, 14) replace the hard-coded 18 and 13/12; dep8 release routing (fpm_generation, generation_priority, measurement_matches_snapshot) with the r6pin library, truth run and pinned lanes on the new generation while the older dep8 evidence and shared decode lanes stay with r4; GENERATION_EVIDENCE_DATES gives only the r6pin snapshot 2026-09-04; SUPPORTING_EVIDENCE_PATHS lists reviewed non-truth files published beside a protocol bundle with the schema's supporting_evidence role under measurements/provenance/ (one sentence added to docs/measurement-protocols.md); GLM campaign, collector, primary and discarded paths; README audit notes and lineage sentence; audit known-gaps lines and a "Listener-window normalization" section in the reconciliation report.
Not imported. The source revision also contains two campaigns by Ethan Shen (glm52-h200-mtp-specdec-validation, with H200 FPM parquets, and dsv4flash-b200-specdec-validation); they are left to their owner to add.
Result. 16 configurations (was 15), 8 history snapshots (7), 28 source parquets and 299,067 rows (25 / 280,365), 202 source measurement files published as 214 artifacts (163 / 173), 25 derived-truth files (18), 15 FPM provenance files (13). validate passes on the rebuilt tree. Outside the dep8 and GLM leaves, only source_revision and dependent hash fields change.
@nv-simonec heads-up and review request. AISimPlayground main moved past the pinned 1a5edbd with two FPM additions, and this PR is the rebuild from the new audited commit c5a1fee: Kimi's DSV4-Pro GB300 dep8 re-measurement with worker pinning (r6pin: new kv-on/kv-off libraries with identical decode rows and re-measured prefill, plus the r6pin truth as derived scores), and my GLM-5.2-NVFP4 GB200 DEP16 campaign (a schema-v6 library for the recommended decode-kvwarm and real-seeded prefill cells, plus the 16/16 disagg truth-run evidence under measurements/).
I followed your conventions as closely as I could read them from the importer: r6pin is a new dep8 generation with r4 moving to history under its existing snapshot id, the derived-truth count is now a constant (18 to 25) in importer and validator, and the GLM campaign gets its own campaign, collector and evidence entries. Rebuilt with pyarrow 25.0.0 and validated locally: 18 configurations, 8 history snapshots, 30 source parquets, 315,158 rows. Outside the dep8 and GLM leaves only provenance fields change.
Two calls I would like from you: (1) the dsv4pro campaign evidence_date stays 2026-09-03 and a small per-generation override gives only the r6pin snapshot 2026-09-04, while the campaign protocol id gained r6pin and therefore shows on the tep8 manifests too; (2) the GLM leaf's truth is derived tables and samples since the raw streams stay on Lustre. The description above lists the other choices worth a look. Happy to adjust anything to your preference.
Thanks for the PR!
(1) the dsv4pro campaign evidence_date stays 2026-09-03 and a small per-generation override gives only the r6pin snapshot 2026-09-04, while the campaign protocol id gained r6pin and therefore shows on the tep8 manifests too
I unified the various protocols into forward-pass-measurement-v1 and forward-pass-record-v1 in https://huggingface.co/datasets/nvidia/aisimulate-fpm-dataset/discussions/4. Now the protocol is experiment-detail-agnostic, without the suffices like r4, r6, etc. Please see forward-pass-measurement-v1.schema.json and forward-pass-record-v1.schema.json for details. (In short, forward-pass-measurement-v1 contains all details like dp rank, request queues, etc.; while forward-pass-record-v1 only contains bare minimal information including "phase", "batch_size", "total_prefill_tokens", "total_kv_read_tokens", "truth_latency_ms" needed by FPM modeling.
(2) the GLM leaf's truth is derived tables and samples since the raw streams stay on Lustre.
Shall we download and upload the data collected on Lustre 👀?
@nv-simonec rebased onto main 1620e9f (your #3 to #6) and force-pushed; the description above is updated. The new contract is followed: the DSV4-Pro campaign keeps forward-pass-record-v1, the r6pin derived truth is flat under measurements/ with the r6pin windows/configs under measurements/provenance/, the two pinned lanes' resolved configs are FPM provenance linked from the new kv-on/kv-off metadata, and your dtype correction applies to the new r6pin libraries. The per-generation evidence date for r6pin is kept, so no existing snapshot is re-dated. Your lane-config predicate fix on main made mine unnecessary.
The GLM-5.2 GB200 DEP16 leaf is back in, with a new library (AISimPlayground MR !20): the decode cell was re-collected on the full paired batch axis (68 rows with every capture size and its +1 neighbour, 1,025 points) and the prefill cell sits on every 8,192-token chunk boundary (7,204 points), so aisimulate can consume it directly; the earlier 12-row pair is discarded with a reason. Its dtype columns already carry the runtime labels, so no correction policy is declared for it.
On the Lustre streams: yes. The GLM bundle is supporting evidence for now (bucket-summary-v1; bucketed tables, samples, aiperf exports). I am preparing the raw per-rank FPM streams of the 16/16 AgentX truth run (16 prefill + 16 decode ranks, about 3.1 M records, 60 to 95 MB gzipped); they will land in AISimPlayground first and then come as a separate PR flipping this bundle to forward-pass-measurement-v1. Two questions for that PR: do you prefer one .jsonl.gz per rank or one per engine role, and how does FPM Gym treat a DP engine's stream, given the "rank 0 as representative record" rule? Our scored truth is the per-step maximum over the 16 ranks joined by receive timestamp, which the iteration_groups representation would express directly.
@nv-simonec the GLM ground truth is in: AISimPlayground MR !21 landed the two raw forward-pass-measurement-v1 observation streams of the 16/16 disaggregated truth run (one per engine role, 16 DP ranks each: 3,056,367 decode-engine and 96,819 prefill-engine records, 67.9 MB and 1.5 MB gzipped), so this PR is re-pinned to cc3524a and the GLM bundle is now a protocol bundle with the two streams flat under measurements/ as truth. The bucketed tables, reports, record samples and aiperf exports moved to measurements/provenance/ as supporting_evidence; to do that I added a small SUPPORTING_EVIDENCE_PATHS table (reviewed non-truth files published beside a protocol bundle, using the role your schema already defines and its provenance path rule). Happy to drop them instead if you prefer protocol bundles to carry truth only.
Two questions from my previous comment still stand for how FPM Gym should read this: the streams are one file per engine role in dp_rank order (the Kimi-K3 and MiniMax convention), and the campaign's step-level truth is the per-step maximum over the 16 ranks joined by the listener receive timestamp (_recv_ts, outside the schema), which the rank-0 representative rule would not reproduce for an attention-DP engine.
Update from PR #7, including commit 5306df8:
- The canonical
forward-pass-measurement-v1iteration retains every contributing rank in authoritativerank_measurements, ordered bydp_rank. - It caches
max_rank_wall_timeas the default scoring latency. Consumers may derive another policy from the retained ranks. - Listener grouping is now
listener_receive_time_window, not exact_recv_tsequality. It records raw timestamps, clock-group corrections, corrected window start/size, and the distinct-rank rule. - This represents the GLM DEP16 method: correct the four node clocks, then cluster distinct ranks in an 8 ms half-open window.
PR #7 enables this representation, but does not materialize synchronized iterations for this PR. The two GLM files here are still physical rank-event streams concatenated by dp_rank.
Actions still needed for this PR:
- Rebase after PR #7 lands.
- During dataset import, reconstruct the prefill and decode streams with the campaign's documented clock correction and 8 ms grouping.
- Emit canonical iteration JSONL containing
expected_dp_ranks,complete, sortedrank_measurements, grouping provenance, andmax_rank_wall_time(nullfor incomplete groups). - Make the canonical files the evaluation truth and retain the raw streams as source/supporting evidence, then regenerate manifest IDs, hashes, catalogs, and validation reports.
- Correct or clarify the campaign truth wording: the checked-in scorer currently reads
wall_med, while the dataset/Gym default selected here is maximum.
@nv-simonec done as requested; rebased onto 32725a9 (#7), re-pinned to AISimPlayground 731947a and force-pushed. The description above is rewritten for the canonical form.
- Rebased after #7.
2./3. The importer now reconstructs both GLM DEP16 streams at import (LISTENER_WINDOW_NORMALIZATIONS, a port of the campaign'struth_table.py): per-node clock groups, offsets from long busy steps shared in lockstep, then half-open 8 ms windows with one record perdp_rank, emitted assynchronizedIterationrecords withproducer,collector, the fulllistener_receive_time_windowgrouping,expected_dp_ranks,complete,max_rank_wall_timeandrank_measurementssorted by rank with_recv_tsretained. Prefill: 96,819 records into 67,220 iterations (182 complete). Decode: 3,056,367 records into 280,194 iterations (91,163 complete). One caveat: on the decode engine no busy step exceeds 150 ms, so the offsets are declared zero rather than estimated, and every decode record says so. The reconstruction matches the campaign's own tool on both streams. - The canonical files are the truth; the raw streams, their provenance record and the summaries are
supporting_evidenceundermeasurements/provenance/. The validator checks each canonical file against the schema's semantic constraints and requires the rank-record total to equal the retained raw stream's record count. - Wording corrected on both sides (AISimPlayground MR !22 is in the pin): the scorer's measured step time is the median over busy ranks, the prediction side is the maximum; on complete decode steps the maximum exceeds the median by about 1 % at the median (p90 1.7 %, p99 8 %).
Two calls for you: the raw streams and the ten per-node/per-tier sample files are now redundant with the canonical files (streams kept per your item 4 and used by the validator's cross-check; the samples can go if you read "redundant copies of rank streams" strictly), and SUPPORTING_EVIDENCE_PATHS is the small table that lets a protocol bundle carry such reviewed non-truth files under provenance/.
Merging now. A focused follow-up PR will remove redundant sample files and make listener-window timestamp arithmetic exact for the Gym adapter.