--- pretty_name: AISimulate FPM Dataset configs: - config_name: catalog default: true data_files: - split: configurations path: catalog/index.parquet - config_name: fpm data_files: - split: files path: catalog/fpm.parquet - config_name: measurements data_files: - split: files path: catalog/measurements.parquet license: apache-2.0 --- # AISimulate FPM Dataset Forward Pass Model (FPM) libraries and independent latency measurements. This private repository is the input-data source for FPM Gym. ## Layout ```text data/ --///// manifest.json # flat metadata for the promoted snapshot fpm/ # optional self-benchmark FPM libraries and sidecars provenance/ # FPM producer configurations measurements/ # live or independent measurement evidence manifest.json # manifest role: truth or derived_truth provenance/ # configuration, window, and manifest helpers history/--///// aisim-commit-/ manifest.json fpm/ measurements/ catalog/ # global discovery indexes schemas/ # current contracts history/ # older manifest/catalog contracts used by migration audit/ scripts/manage_dataset.py ``` Model directories use the exact case-sensitive Hugging Face model ID with `/` replaced by `--`. Example active leaf: ```text data/MiniMaxAI--MiniMax-M2.7/h200-sxm/vllm/0.25.1/dep4/ ``` Campaign names never define canonical paths. They remain only in provenance. Multi-configuration source Parquets are split by their exact runtime identity, so a parallelism leaf cannot silently contain rows for another configuration. Usable measurement truth is flat under each `measurements/` directory. Supporting summaries, configurations, windows, and manifests retain useful source grouping under `measurements/provenance/`. Only configurations explicitly linked to an FPM artifact are stored under `fpm/provenance/`. Each leaf `manifest.json` is flat. Model, hardware, framework, parallelism, AISim commit, and collector fields apply directly to the promoted data. There are no `identity` or `current` wrappers. Its optional `history` list points to older immutable snapshots. Measurement-only configurations use `fpm: []` and omit the `fpm/` directory. They must contain validated measurement truth; an entirely empty or summary-only configuration is not admitted through this path. The catalog records `fpm_file_count: 0` independently of measurement availability. Missing predictor data does not invalidate a measured point. Four AgentX captures are published at standard configuration leaves. See [AgentX collection and processing](docs/agentx-fpm-20260911.md) for their exact identities, grouped representation, provenance, and reproducible checks. ## Kimi K3 DCP self-benchmark The [TP8+DCP8 self-benchmark configuration](docs/kimi-k3-self-benchmark-20260916.md) adds 669 primary balanced points, 14 synthetic-attention boundary comparator points, and 56 nonuniform prefill points preserved as calibration provenance. The `aic_fpm_forward_perf` v6 schema records optional DCP separately from independent CP; this configuration explicitly records `dcp: 8`. Consumers must preserve that identity or report DCP as unsupported. Existing v6 artifacts and the older Kimi TEP8 configuration remain unchanged. Sidecars and catalogs record `collection_method` and `source_kind`; these calibration inputs are not independent measurement truth. ## Artifact and measurement contracts FPM Parquets are predictor inputs. The existing AIC libraries use schema name `aic_fpm_forward_perf` and schema version 6; they do not have a measurement protocol. Their logical schema is [`schemas/aisim-fpm-forward-v6.schema.json`](schemas/aisim-fpm-forward-v6.schema.json). The pinned DSV4-Flash, DSV4-Pro, Kimi-K3, and GLM source Parquets contain reviewed dtype annotation errors. During publication, a strict campaign- and model-scoped policy corrects only these declared fields: - DSV4-Flash: MoE `fp8` → `w4a8_mxfp4_mxfp8`; FMHA `fp8` → `bfloat16`. - DSV4-Pro: MoE `fp8_block` → `w4a8_mxfp4_mxfp8`. - Kimi-K3: MoE `fp8_block` → `w4a16_mxfp4`. - GLM-5.2-NVFP4: FMHA `fp8` → `bfloat16`. DSV4-Flash and GLM FMHA resolution is also corrected from `checkpoint_native` to `aic_data_fallback_from_fp8`. The importer fails if the campaign, model, or source labels differ from the reviewed policy. It preserves GEMM, weight quantization, latency, workload kind, batch size, token coordinates, row order, and row count. Each affected sidecar records the `runtime-dtype-annotation-v1` policy and exact source-to-published labels. Usable measurement bundles declare one of two semantic protocols: - `forward-pass-measurement-v1`: - Normalizes exact scheduler observations into one synchronized iteration. - Retains one measurement for every contributing rank and caches `max_rank_wall_time` as the default complete-iteration truth. Consumers may derive another latency policy from the authoritative rank measurements. - Distinguishes two source kinds: - `self_benchmark`: a benchmark producer groups ranks by `benchmark_id`. - `rank_event_stream`: rank-local events are either evaluated directly for one rank, joined using a declared corrected receive-time window, or supplied as pre-grouped rank lists with uploader-provided membership and provenance. - `producer` identifies the component that creates rank-local FPM records. Optional `collector` provenance identifies a separate capture component; it is required when grouping by `listener_receive_time_window`. - `grouping` records the membership method and its authority. Consumers do not implicitly join ranks by local counters, file position, or timestamp equality. With `pre_grouped`, membership is already stored; inferred alignments remain labeled `derived` and retain their audit provenance. - Schema: [`forward-pass-measurement-v1.schema.json`](schemas/forward-pass-measurement-v1.schema.json). - `forward-pass-record-v1` contains a reduced, self-contained point: scheduled batch/token coordinates plus observed latency. The current DSV4-Pro reduced measurement records use this protocol because their raw streams are external. See [`schemas/forward-pass-record-v1.schema.json`](schemas/forward-pass-record-v1.schema.json). Workload kind (`prefill`, `decode`, `mixed`, or `empty`) is derived per record; it is not part of the protocol ID. Producer names, traffic sources, collection topology, and release labels such as `r4`, `r6pin`, and `r16` stay in provenance. Summary-only artifacts cannot recreate predictor input. They set `measurement_protocol_id: null` and use `evidence_format_id`, such as `wave-summary-v1` or `bucket-summary-v1`. They remain visible as supporting evidence but are not usable measurements. See [measurement protocols](docs/measurement-protocols.md) for field meanings, serialization rules, and normalization behavior. ## Promotion and history One reviewed snapshot is promoted for each full configuration. The identity includes model revision, system, framework version, all parallel dimensions, weight quantization, and KV-cache dtype. - Explicit release lineage decides between generations such as R17/R16 and DSV4-Pro r9/r8/r3 and DEP8 r6pin/r4. - A later reviewed campaign supersedes an earlier campaign only where their full configuration identities overlap. - Legitimate variants from one promoted snapshot, such as KV-warm and KV-off, remain together. - Superseded FPM and measurement artifacts move together under `data/history/`. - A valid recorded producer SHA uses `aisim-commit-` as the snapshot directory. Missing producer SHAs use `aisim-commit-unknown-` and keep `aisim_commit: null`; the importer never invents a commit. The top-level manifest records `collected_by` and how that attribution was derived. `aisim_commit` describes the producer code when known. `provenance.source_revision` separately records the audited AISimPlayground source commit. ## FPM Gym evaluation This repository stores source evidence, not evaluation output. FPM Gym: - Parses the measurement artifacts for a selected configuration snapshot. - Runs AISim against the matching FPM artifact at evaluation time. - Derives measurement and prediction states independently. - Applies exclusions and computes coverage and accuracy metrics. No predictions, normalized comparison rows, selector mappings, or accuracy metrics are persisted here. Configuration identity comes from the leaf path and manifest. Workload coordinates come from the FPM and measurement contents. See the [evaluation state policy](docs/evaluation-state-policy.md) for the measurement and prediction state tables, exclusion rules, and metric denominators. The trend page can include history by following manifest history references or filtering `catalog/index.parquet` by `snapshot_status`. ## Publication policy - Consumers must pin a full Hugging Face commit SHA. - Discovery starts from `catalog/index.json` or an active leaf manifest. - Consumers verify every recorded SHA-256 before loading data. - FPM Gym records the pinned dataset commit and AISim commit for each evaluation. - Missing or invalid measurement evidence remains visible as a gap; it is never scored. - Raw logs, reproduction scripts, reports, redundant copies of rank streams, and wholesale prediction directories are not mirrored. The repository must remain private. Retained provenance contains internal paths and network metadata, and the source has no approved redistribution license. ## Audit notes - DSV4-Pro raw serving streams are external; normalized score truth and its source hash are retained. - Usable measurement truth files are flat under `measurements/`; summaries, serving configs, window maps, and helper manifests are under `measurements/provenance/`. - DSV4-Pro benchmark-lane configs are linked from the matching FPM metadata and stored under `fpm/provenance/`. - DSV4-Pro TEP8 r9 is active. Diagnostic r8 and superseded r3 evidence are in history. - DSV4-Pro DEP8 r6pin (worker-pinned prefill lanes, decode rows identical to r4) is active; the r4 library and its evidence are in history. - MiniMax H200 R17 is active for `dep4` and `pure-tp4`; R16 is history. The later static snapshot is active for H200 `tep4`. - MiniMax B200 repaired evidence supersedes the earlier decode-only E9 snapshot. - MiniMax B200 `dep4/c4` is quarantined, and `dep4` prefill has no retained measurement evidence. - GLM and MiniMax H200 static anchor inputs are incomplete. - GLM-5.2 GB200 DEP16 (`data/nvidia--GLM-5.2-NVFP4/gb200/vllm/0.28.0/dep16/`) publishes the paired-rows decode and chunk-boundary prefill self-benchmark library with canonical `forward-pass-measurement-v1` iterations of the 16/16 disaggregated truth run, one file per engine role, reconstructed at import from the raw rank-event streams with the campaign's per-node clock correction (estimated for prefill; declared zero for decode, where no shared step exceeds 150 ms) and 8 ms half-open window (`listener_receive_time_window`, exact decimal timestamp arithmetic, every rank record retained, `max_rank_wall_time` cached). The raw streams, bucketed tables, reports and aiperf exports are supporting evidence under `measurements/provenance/`. The campaign's published fidelity figures used the median wall over busy ranks; on complete decode steps the maximum exceeds the median by about 1% at the median (1.7% at p90, 8% at p99). Decode is 3.6% and prefill 24.0% step-level MAPE against the 16/16 disaggregated truth (17.2% when all 16 ranks prefill); vLLM 0.28.0 is not a curated aisimulate version slot. Its source dtype labels already carry the runtime identity, so no publish-time correction applies. The superseded 12-row decode pair is discarded. See [source mapping](audit/source-mapping.md), [reconciliation](audit/reconciliation-report.md), and [validation output](audit/validation-report.txt). ## Developer guide: contribute a dataset Submit additions as a Hugging Face dataset pull request against current `main`. Include the data, sidecars, manifests, regenerated catalogs, and reproducible validation evidence in the same PR. Use the contracts below as the checklist; copying an existing directory does not establish schema compliance. ### Choose the artifact contract | Artifact | Required contract | Key requirements | | --- | --- | --- | | Configuration leaf `manifest.json` | [Configuration manifest v4](schemas/configuration-manifest-v4.schema.json) | Flat runtime identity, producer revision/status, collector attribution, FPM and measurement references, source provenance, and history. | | FPM Parquet and adjacent `.metadata.json` | [FPM v6 logical schema](schemas/aisim-fpm-forward-v6.schema.json) | Exact query identity, workload coordinates, field types, positive finite `latency_ms`, row count, and matching digest/sidecar. | | Reviewed DeepSeek V4.1 schema-7 FPM | [Schema-7 admission policy](scripts/dsv41.py) and [profile guide](docs/dsv41-pr304.md) | Preserve schema 7 and source-qualified execution identity, including `full` versus `decoder_bounded`; do not relabel as v6. | | `measurements/manifest.json` | [Measurement manifest v5](schemas/measurement-manifest-v5.schema.json) | Protocol ID, file roles, repository-relative paths, published/source hashes, source revision, and grouping provenance where required. | | Canonical measurement storage | [Arrow schemas](scripts/measurement_storage.py) and [storage v1 descriptor](schemas/measurement-storage-v1.json) | Parquet bulk records, explicit storage schema/version, logical row count, ordering index, and collection-run bindings. | | Full scheduler/rank observations | [Forward-pass measurement v1](schemas/forward-pass-measurement-v1.schema.json) | Explicit iteration membership, all contributing rank observations, and declared timing/grouping policy. | | Reduced measured points | [Forward-pass record v1](schemas/forward-pass-record-v1.schema.json) | Batch/token coordinates and finite positive `truth_latency_ms`, with source provenance. | | Discovery indexes | [Catalog index v6](schemas/catalog-index-v6.schema.json) | Consistent configuration/history references, counts, and generated JSON/Parquet catalogs. | The JSON manifest schemas define required fields and allowed values. The FPM v6 file is a logical Arrow contract, not a JSON-record schema. Use the dataset validator for physical types, sidecars, and cross-file checks. Schema versions do not make campaign contracts interchangeable. DeepSeek V4.1 uses its own admission code; Inkling uses the [Inkling schema-7 extension](schemas/aisim-fpm-forward-v7-inkling.schema.json) and [admission policy](scripts/inkling_six_run.py). GLM uses the [GLM admission policy](scripts/glm53flash.py). New schemas or protocols need corresponding validation support and tests in the same PR. Current contributions use configuration v4, measurement v5, and catalog v6. Older manifest/catalog contracts live in [`schemas/history/`](schemas/history/) for legacy validation and migration. New contributions must use the current contracts in `schemas/`. FPM variants and measurement protocols still in use remain in `schemas/`; their version numbers describe separate contracts, not a shared upgrade sequence. ### Prepare the contribution 1. Start from the latest dataset `main` with actual data files downloaded, not Git LFS pointer files. Pin the source repository/archive revision and record source SHA-256 digests before transforming data. 2. Place each exact configuration under `data/--/////`. Preserve case-sensitive model IDs, model revision, topology, quantization, KV-cache dtype, and optional DCP. Keep distinct runtime identities separate. 3. Put predictor libraries and sidecars under `fpm/`. Put usable truth directly under `measurements/`; put its supporting files under `measurements/provenance/`. Calibration inputs must retain their collection method and must not be presented as independent measurement truth. 4. Populate both manifests using the linked schemas. For measurement-only contributions, use `fpm: []`, omit `fpm/`, and include validated truth. Summary-only evidence uses a null protocol and an evidence format ID. Calibration-only bundles without measured truth retain `files: []` and `collection_runs: []`; do not invent measurement runs. 5. Use an existing matching ingestion path, or add a focused importer and validation tests for a new source format. Existing examples include [AgentX ingestion](scripts/ingest_agentx.py) and [DSV4.1 ingestion](scripts/ingest_dsv41.py). These importers are source-specific; the base `import --replace` command is not a general contribution command. 6. Regenerate catalogs through the ingestion code and `write_catalogs` in [manage_dataset.py](scripts/manage_dataset.py). Preserve unrelated configurations. When replacing a promoted snapshot, move its FPM and measurement artifacts together into history and update references. For rank measurements, follow [measurement protocols](docs/measurement-protocols.md): local counters and file order are not implicit cross-rank join keys. Declare producer or derived grouping authority, expected ranks, and hashed grouping provenance. Retain incomplete observations as evidence without labeling them complete-iteration truth. ### Measurement storage and collection metadata - Publish bulk iterations, reduced records, iteration indexes, and request metrics through the shared storage writer: Zstd level 3, 16,384-row groups. Manifests/settings/provenance remain JSON. Preserve null versus zero, int64 nanoseconds, rank membership, completeness, and source ordering. - Each measurement file declares `format`, `storage_schema`, `logical_row_count`, `ordering_index`, hashes, and immutable source references. Indexes use `record_file_id` and one-based `record_id`, not physical row offsets. - Bind truth records to `collection_runs` using disjoint half-open logical ranges; a null range covers the whole file. Unattributed truth belongs to an explicit unknown run. Record collection type, preset, collector, workload, replay mode, and serving layout from evidence; names do not establish them. - Preserve request-level timing boundaries and outcomes. Do not reconstruct request metrics from percentiles or forward-pass iterations. See [measurement v5 and collection runs](#measurement-v5-migration-and-collection-runs) for naming, metadata, and storage details. ### Validate and request review Run from the dataset root after downloading the referenced payloads: ```bash uv run --with pyarrow==25.0.1 --with jsonschema --with ijson \ python scripts/manage_dataset.py validate uv run --python 3.12 --with pyarrow==25.0.1 --with jsonschema --with ijson \ --with pytest --with numba python -m pytest scripts git diff --check ``` Include these items in the PR description: - **Scope:** added/replaced configurations, file and row counts, schema/protocol versions, and whether this is new collection, transformation, or migration. - **Provenance:** immutable source revision, producer identity, collection command/configuration, model revision, runtime/package versions, and licenses. - **Integrity:** validation/test commands and results, source/published hashes, and any intentional transformations. For unchanged migrations, compare payloads byte-for-byte against the pinned source. - **Qualification:** calibration versus heldout/independent evidence, supported workloads, missing measurements, exclusions, and known collection failures. - **Consumers:** affected AISimulate references and the full dataset commit SHA and file hashes they will pin. Refresh pins after changing metadata. The shared validator dispatches each imported FPM artifact to its admission policy and requires a unique matching validated snapshot. Hashes alone do not establish model, hardware, parallelism, quantization, or execution identity. When adding an importer, test mismatched identities and remove/re-import catalog stability, not only successful ingestion. Reviewers check the exact submitted commit. Passing storage validation establishes integrity; it does not establish new GPU accuracy or remove existing quarantine and qualification limits. ### Lessons from dataset PR reviews - Keep `aisim_commit` for the producer code and `provenance.source_revision` for the source archive. Preserve original producer identities; use null with `unknown` or `mixed` when no single producer SHA is established. See [PR #11](https://huggingface.co/datasets/nvidia/aisimulate-fpm-dataset/discussions/11). - Removing prediction/error metrics must preserve collection evidence: image digests, known checkpoint revisions, package versions, attempt-selection policy, and runtime qualification. A source-sidecar hash alone does not retain these fields. See [PR #12](https://huggingface.co/datasets/nvidia/aisimulate-fpm-dataset/discussions/12). - If another dataset PR merges first, update from `main`, regenerate combined catalogs, and rerun validation. Resolve catalog conflicts by retaining both contributions; verify uploaded files and consumer hashes at the reviewed SHA. ## Validate or refresh Validation is source-independent: ```bash uv run --with pyarrow==25.0.1 --with jsonschema --with ijson python scripts/manage_dataset.py validate ``` Rebuild from the audited internal checkout: ```bash uv run --with pyarrow==25.0.1 --with jsonschema --with ijson python scripts/manage_dataset.py import \ --source-root /path/to/AISimPlayground/playground \ --replace ``` The importer rejects any source revision other than the pinned audited commit. It also refuses to overwrite externally contributed measurement-only leaves. Rebuild the base import into a fresh destination, then reapply reviewed external ingestions. This prevents a Playground refresh from silently deleting AgentX data. For the pinned AgentX package, `scripts/ingest_agentx.py` performs a collision- checked migration and regenerates the standard catalogs. Grouped-list reading and canonical normalization live in `scripts/pre_grouped.py`; its validator uses stored membership and does not rerun alignment or impose a receive-time window. ## DeepSeek V4.1 GB200/H200/B200 TP4 full and decoder bounded [The PR #304 profile guide](docs/dsv41-pr304.md) covers six independently collected TP4 profiles: GB200, H200 and B200, each `full` and `decoder_bounded`. Each has 145 source-qualified native calibration rows and a separate 38-geometry heldout set. The guide lists coverage, heldout accuracy, replay entry points and limitations, including the 12 geometries excluded by the bounded aggregate API and the retained B200 full prefill errors. ### Inkling GB300 calibration import The `inkling-gb300-three-run-20261007` campaign covers `thinkingmachines/Inkling-NVFP4` revision `42a75a99a40eb2ba1e0717db6357a0bf15205044` on GB300 with vLLM 0.27.0, NVFP4 weights and BF16 attention/KV. This publication contains pure TP4, pure TP8 and DEP4, with three independent full-grid launches per phase (prefill and decode), for 18 launches in total. DEP8 is deferred and excluded from this publication. The schema-7 estimator takes the maximum latency across DP ranks within each launch, then the median across all three launches. The original source measurement is excluded from the estimator; source comparability remains required. The coefficient of variation (CV) must be at most 5%. Temporal split-half drift is `not_assessed` for three runs. Held coordinates and their observations remain in the proof populations and are excluded from the published FPM table. The same importer and validator also support the legacy `inkling-gb300-six-run-20261005` policy, with six-launch medians and its additional 5% temporal split-half drift gate. Run `scripts/ingest_inkling_six_run.py export --descriptor --output ` in the pinned producer environment, where the original campaigns are available. The descriptor names `topology`, `evidence_date`, `parquet` (the merged exact-topology canonical pair), optional `metadata`, and `cells`, each containing `repeatability_dir` and its complete `assessment` path. Export reassesses the original campaigns without scheduling GPUs, preserves canonical duplicate selection and all held observations, and creates a portable package. Import locally with `scripts/ingest_inkling_six_run.py import --dataset-root . --package `. Both commands are offline; neither uploads data or changes repository visibility. The public projection explicitly selects nested measurement, graph, runtime and CPU fields. It aliases host/process/thread identity while retaining rank placement and affinity relationships. Original artifact/signature hashes remain historical provenance; projected launch files, population JSON, Parquet and sidecar each receive hashes of their actual new bytes. Original files remain in the campaign archive and are required for the original collector resume/reassessment commands. The portable validator independently checks raw internal timing reductions, the declared three- or six-run estimator and gates, row/population bindings, zero-KV duplicate selection, and evidence-backed workload imbalance holds. Latency imbalance alone does not exclude a point. The proof inventory must include exactly one prefill and one decode population, even when an entire phase is held. Export rederives formal KV provenance with the canonical first-launch aggregator and binds every native seed to the declared launch observations and rank preparation. Canonical legacy prefill keeps formal `n/a`; its native `not_applicable` or `real_prefix` evidence remains separate. This campaign does not claim either generic KV warm-up skip exemption. The validator rejects extra fields recursively, and the Inkling metadata, import proof and launch JSON paths use Git LFS. These are calibration inputs with scope `measurement_repeatability_only`; serving accuracy is `not_assessed`. Configuration manifests use v4 and measurement manifests use v5, with no truth files, `collection_runs: []`, and a null protocol. Import appends to catalog v6 without rerunning the historical measurement migration; campaign identity remains in the checked portable proof rather than normalized configuration provenance. Holdout, replay, and serving accuracy reports remain outside this dataset. The format extension is documented in [the Inkling schema-7 contract](schemas/aisim-fpm-forward-v7-inkling.schema.json). The original campaign publication and qualification used AISimulate `7d2c0a692ced4a433211819705958c04463dc146`. Ordinary FPM loading is now verified with AISimulate [PR #379](https://github.com/ai-dynamo/aisimulate/pull/379) at commit `20e3eb9d52325a3e5a24e34c54dd09da130346fd`: the loader consumes each aggregated Parquet/metadata pair without campaign files, preserves all 10,264 retained latencies, and returns unchanged results for 10,849 checked exact and interpolated queries. This is a loader and numerical-regression check, not an independent accuracy measurement. Use that commit, or a version containing its loader change; compatibility with earlier AISimulate versions is not asserted. [Version provenance](docs/inkling-three-profile-version-provenance-20261011.json) keeps planning, actual collection, original qualification, and later consumer validation revisions separate. | Profile | Retained prefill coordinates | Retained decode coordinates | Held coordinates | | --- | ---: | ---: | ---: | | TP4 | 2,821 | 788 | 92 | | TP8 | 2,740 | 938 | 180 | | DEP4 | 2,429 | 548 | 132 | The archived holdout evaluation is not a full serving-accuracy pass: TP8 decode has 35.3521% p95 relative error against a 20% target, and TP4 and TP8 prefill each have one unsupported holdout. These historical results have not been erased by later loader or query-coverage improvements. The 1,560-request CPU replay was attempted for TP4, TP8 and DEP4; all three are incomplete. Each exits with code 1 at its first observed unsupported prefill query with batch size 1, 1,536 prefill tokens and 253,952 KV-read tokens, without hitting a resource limit. Before stopping, TP4 resolved 1,541 measured and 427,477 interpolated queries; TP8 resolved 1,577 and 433,460; DEP4 resolved 2,049 and 473,890. These counts are native lookup resolutions, not completed requests; the failed CLI does not report a completed-request count, and later coverage gaps remain unassessed. Nearby measured points include (prefill=1,024, KV=254,944) and (prefill=2,048, KV=253,920), but they do not provide the same-batch, two-sided support required by the current token-first or KV-first direct interpolation. No extrapolation, new collection or interpolation change was used to turn these attempts into a pass. A separate frozen 24-request serial play completes prediction with zero unsupported queries on all three profiles. Each profile also completed three independently reviewed matched-serving repetitions (72 requests), with protocol, runtime and memory checks. TP4 retained its first-repetition TTFT p95 error of 20.37%; DEP4 retained all three TTFT p95 errors of 20.87–21.02%, above the unchanged 20% reference threshold. TP8 TTFT p95 error was 12.21–16.05% and TPOT p95 error was 4.46–4.52%. DEP4 had observed attention-rank imbalance; no accuracy samples were excluded. TP8 ran on AWS GB300 hosts while TP4 and DEP4 ran on OCI GB300 hosts, with a recorded driver difference; performance equivalence is not assumed. These campaign-local diagnostics use reconstructed prompts and include original trace idle gaps, so they do not establish canonical Stage 6 qualification, saturated throughput or broad serving accuracy. [The validation record](https://huggingface.co/datasets/nvidia/aisimulate-fpm-dataset/discussions/17) preserves the repetitions and limitations separately from these calibration tables. The table metadata remains `serving_accuracy: not_assessed`; the accepted larger-replay gap and original holdout results remain unchanged. ## Measurement v5 migration and collection runs Catalog v6 points to configuration v4 and measurement v5 manifests. Measurement protocols keep their scientific meaning; each file separately declares `format`, `storage_schema` (including version), `logical_row_count`, `ordering_index`, SHA-256, and immutable source references. Readers of older pinned revisions must retain legacy JSON/CSV support. Bulk iterations, reduced records, iteration indexes, and request metrics use the explicit Arrow schemas in `scripts/measurement_storage.py`: Zstd level 3, 16,384-row row groups, and bounded conversion batches. Grouped iterations retain rank lists and their original order, completeness and grouping authority. Ungrouped observations remain explicitly ungrouped. Nanoseconds use int64; latencies use float64; missing values are distinct from zero. Index references use `record_file_id` plus one-based `record_id`, not physical Parquet row positions. JSON remains canonical for manifests, collection settings and provenance. Original compressed raw captures remain evidence, not a second active truth representation. `collection_runs` separates collection type (`self_benchmark`, `static_serving`, `trace_replay`, `unknown`), benchmark preset, collector, replay mode, dataset, workload, and serving topology. AgentX is recorded through an explicit preset; a dataset or filename does not imply its replay semantics. Numeric/string unknowns are null; categorical unknowns are `unknown`. `not_applicable` is distinct. `collection_runs[].serving.layout` records deployment independently of collection type: `agg` uses the same workers for prefill and decode; `disagg` separates the two roles. Both layouts apply to self-benchmarks when supported by capture evidence. Use `unknown` when deployment information is missing; use `not_applicable` only for captures explicitly outside a serving deployment. Starting a real engine or recording prefill/decode phases alone does not establish the deployment layout. `worker_roles` retains `prefill`, `decode`, `aggregated`, or `unknown`; `topology` independently records a known topology such as `1P1D`. Each run binds truth file IDs to disjoint half-open logical record ranges. A null range means the whole file. Every unattributed truth record belongs to an explicit unknown run; empty bundles invent no runs. Configuration and evaluator selection must restrict metadata to the truth actually used. Source references declare which fields they support. Canonical paths and labels omit campaign rounds/scheduler IDs; unchanged upstream evidence and immutable source references preserve original names. Truth Parquet filenames include verified concurrency settings, for example `iterations-c16-decode.parquet`. Multiple values with the same recorded unit use their minimum and maximum, such as `iterations-c16-128-decode.parquet`; this does not imply every intermediate value was measured. The manifest retains the full set of settings. A range with unattributed records adds `-unknown`, for example `iterations-c16-128-unknown-decode.parquet`. Missing settings, unknown or conflicting units, and a single known setting mixed with unknown records retain directory-local numbers such as `iterations-01.parquet` or `records-01.parquet`. The phase label preserves the source filename's distinction. Existing numbers stay fixed; new numbered files append after the highest number. Provenance retains descriptive source basenames and explicit concurrency folders, for example `provenance/c64/agx_windows_decode.tsv`, rather than `asset-/evidence.tsv`. Same-name assets receive numeric suffixes without overwriting existing captures; converted supporting indexes use names such as `iteration_index.parquet`. The manifest's `measurement_file_id` remains the stable identity. Collection metadata and `ordering_index` are authoritative; names never fill missing metadata or determine evaluation order. Renaming does not rewrite file contents. Immutable source URLs retain their original names, and original archives retain their format. Request Parquets record run/request identity, original source record, stage, outcome, observed input/output tokens, request-start offset, TTFT, TPOT and E2E. `timing_boundaries_compatible` gates TPOT derivation. The matched Kimi trace and recovered AIPerf exports supply chart populations; warmup, profiling, drain and outcomes remain explicit. Missing request metrics are never reconstructed from percentiles or FPM iterations. The migration is repeatable and all active importers normalize at the shared catalog publication boundary. Validate or rebuild the complete audit: ```bash uv run --with pyarrow==25.0.1 --with jsonschema --with ijson \ python scripts/migrate_measurements.py --root /path/to/migrated-dataset \ --legacy-root /path/to/legacy-dataset --playground /path/to/AISimPlayground ``` Run this on an isolated checkout. The legacy source is never modified. [The migration audit](audit/measurement-migration-v5.json) covers all 39 snapshots, asset renames, source hash verification, recovered runs, unknown fields and chart availability. Static serving sweeps are distinct from trace replay. AIPerf summaries without verified request/capture alignment remain unresolved. Deploy compatible consumers first, evaluate an immutable migrated HF branch, then promote HF main after those consumers are available. Historical evaluations remain unchanged. ### Request evidence recovery The request evidence recovery inputs in [`audit/request-metrics-recovery-inputs.json`](audit/request-metrics-recovery-inputs.json) pin 19 recovered AIPerf captures to existing measurement truth. Capture matching uses the already-published summary hash or byte-identical uncompressed FPM streams. Original metric-only JSONL exports are losslessly compressed under opaque run paths; AIPerf summaries remain JSON. These archives retain integer nanosecond timestamps and original source-record order. They contain benchmark metadata and metrics, not prompts or generated responses. Their original host, path, file hashes and timezone are recorded in the recovery inputs. These are NVIDIA measurement artifacts from the same collections already represented in this dataset. They were retrieved from ComputeLab, Slurm and Kubernetes PVC storage on 2026-10-09. The AgentX collector source is [SemiAnalysisAI/aiperf at 754356e9](https://github.com/SemiAnalysisAI/aiperf/tree/754356e9a39acc6cc6afb242d123bb57c3fb6f75); Playground collection evidence is pinned in the existing manifests and recovery audit. Request timing is client-visible; FPM truth remains forward-pass latency. The source audit also preserves the original H200 benchmark logs and their file inventory under `audit/request-metrics-source-evidence/`. All 84 benchmark logs disable per-request result export. Their random-dataset arguments conflict with ShareGPT in the copied driver documentation, so dataset identity remains unknown for these captures. The [request recovery audit](audit/request-metrics-recovery.json) covers all 39 current and historical manifests. It distinguishes exports recovered from storage, exports disabled by the collector, missing source directories, supporting-only windows and unattributed records. Together with five existing Kimi runs, 24 runs have request metric assets. Historical Gym evaluations remain unchanged until a new evaluation selects this migrated revision. Recover and refresh catalogs with the shared publication command (requires the archived assets in this branch): ```bash python scripts/migrate_measurements.py --root . python scripts/migrate_measurements.py --root . --validate-only python -m pytest scripts/test_recover_request_metrics.py ``` Recovery checks archive hashes, published capture identity, model, benchmark ID, profiling windows and request counts before adding canonical request Parquets. TPOT uses AIPerf `inter_token_latency`, whose end boundary is the last content token; HTTP response-completion latency is not substituted. Naive timestamps are interpreted using the capture's verified timezone. Nanosecond integers are subtracted before converting offsets to milliseconds. Only collection metadata bindings change: GLM GB200 groups require every rank's receive time to fit one profiling window. Historical MiniMax B200 uses the strict interior of its verified source windows because producer logs and old scoring code disagree on endpoint conventions. Boundary records stay unknown. Truth files, predictor inputs, rank membership and evaluation order are unchanged. Recovered E9 exports omit an explicit replay profile and preset; those fields remain unknown rather than inferring AgentX from a directory name. ## GLM-5.3-Flash GB300 TP1/TP2/TP4 (AISimulate PR #324) [The GLM-5.3-Flash guide](docs/glm53flash-gb300-fpm.md) covers nine GB300 calibration tables for `zai-org/GLM-5.3-Flash` (FP8) and `nvidia/GLM-5.3-Flash-NVFP4`. All are collected with **engine-native forward-pass metrics (FPM) on the default serving configuration, with prefix caching on**. Both backends are final. Admission code is `scripts/glm53flash.py`; the source revision is AISimulate `055dcc1c`. The leaves use configuration manifest v4 and calibration-only measurement manifest v5 (`files: []`, `collection_runs: []`); campaign names and job IDs remain only in each leaf's source evidence under `fpm/provenance/`. Admission checks bind model, hardware, parallelism, quantization, and snapshot identity to the reviewed table, and require the configuration, measurement manifest, and catalog entries to agree. Checksums alone do not establish configuration identity. The shared catalog-v6 validator dispatches GLM, Inkling, DSV4.1, and self-benchmark imports to their own admission policies and requires a unique validated snapshot for each imported catalog entry. GLM and Inkling imports use the shared Parquet catalog writer so remove/re-import round trips preserve catalog checksums. - **vLLM: stock v0.31.0** (campaign `glm53flash-vllm-v031-native-fpm`). FP8 TP2/TP4 and NVFP4 TP1/TP2/TP4; FP8 does not fit on one GPU, so TP1 is NVFP4 only. - `latency_ms` is the median of ten repetitions of the engine-native Dynamo `InstrumentedScheduler` FPM `wall_time`, with vLLM's own scheduler running every request. - Prefill inputs are align-4 (prefix and new-token lengths on the 4-token IndexPool grid); this is a collection rule, not a prediction restriction. - **SGLang 0.5.20** (campaign `glm53flash-pr324-native-fpm`). FP8 and NVFP4 at TP2/TP4. `latency_ms` is the median of ten repetitions of the native rank-0 `ForwardPassMetrics` `wall_time` (DeviceTimer), with the radix cache on. Holdout FPM MAPE, prefill / decode (gate 10% per phase; all 18 phases pass): | Backend | FP8 TP2 | FP8 TP4 | NVFP4 TP1 | NVFP4 TP2 | NVFP4 TP4 | | --- | --- | --- | --- | --- | --- | | vLLM 0.31.0 | 2.16% / 1.99% | 1.28% / 1.63% | 2.35% / 1.98% | 3.68% / 1.53% | 2.47% / 1.13% | | SGLang 0.5.20 | 1.57% (n=142) / 1.99% | 2.08% / 1.27% | n/a | 2.44% / 1.79% | 2.19% / 1.13% | SGLang FP8 TP2 is missing two calibration points, which got no radix hits on the stock server; as a result, two of its 144 prefill holdout points are outside the collected domain and are not predicted. Nothing is imputed. The earlier custom-timer tables (`glm53flash-pr324`, `glm53flash-pr324-prefill-graph`), the native-FPM vLLM `0.30.0+glm53tail` tables and the superseded vLLM NVFP4 TP4 table from calibration-decode job 904065 were never merged; the admission policy rejects every one of their hashes. ## Experimental GLM-5.3-Flash B300 FP8 TP4 The expanded table contains **427 prefill and 175 decode cases** for `zai-org/GLM-5.3-Flash` on B300 SXM, stock vLLM 0.31.0, TP4, FP8 weights and sparse KV, prefix caching and Mamba align mode. It retains all 359 original calibration medians unchanged and adds 243 disjoint native measurements around small-input coverage and the observed 64-token execution-mode boundary. The producer and consumer interpolation code are unchanged. Every case has five warmups and ten measurements; a separate launch repeats the new 243-case grid. The original prefill validation failed and is preserved in dataset commit `4dc9dda1ccc5f6b778c0bbea5138b28237082b2f` and the companion integration evidence. After freezing the expanded table, 68 new independent cases were collected: 53/53 prefills are supported with **2.68% average absolute percentage error** (maximum 10.68%), and 15/15 decodes with **1.10%** (maximum 3.69%). Both pass the unchanged 10% per-phase mean-error threshold. On the 45 identical fresh prefill cases supported by both tables, average error falls from 76.70% to 2.50%; the original table also leaves eight fresh prefill cases unsupported. **Admission remains `development`.** Validation covers synthetic align-4 request lengths and finite standalone replay; it does not establish matched serving or cache-reuse accuracy, arbitrary unaligned request lengths, 1M context coverage, other parallel layouts, or a supported DynoSim dependency pair. Source-inclusive two-launch variation is diagnostic, not full canonical fresh-population qualification. Default materialization is blocked; explicit unqualified reproduction is diagnostic only. No valid slow measurements were removed. The leaf keeps calibration-only manifests with empty normalized truth files and collection runs. Its provenance requires both native source jobs and hashes. All nine GB300 profiles retain their original payload bytes and policies. See [integration draft #511](https://github.com/ai-dynamo/aisimulate/pull/511) for complete before/after evidence and the retained original failure. ## GLM-5.3-Flash GB200, B300, H200 and H100 (AISimulate PR #515) 19 calibration tables collected with the GB300 method above (engine-native FPM, default serving configuration, prefix caching on) on GB200, B300, H200 and H100, from AISimulate [PR #515](https://github.com/ai-dynamo/aisimulate/pull/515) at `1a73bcfe`. Campaigns `glm53flash-multigpu-vllm-v031-native-fpm` and `glm53flash-multigpu-sglang-v0520-native-fpm`; admission code is `scripts/glm53flash.py`, which also keeps the GB300 tables at `055dcc1c`. - KV-cache dtype is part of the identity: FP8 on GB200 and B300; BF16 on H200 and H100 for both backends (stock vLLM 0.31.0 and SGLang 0.5.20 do not serve FP8 sparse-MLA KV on SM90). - The vLLM and SGLang images are pinned by digest per CPU architecture (arm64 on GB200, amd64 on B300, H200 and H100). TP is 1, 2, 4 or 8. - Request tokens are seeded ids (seed 53) for both backends, recorded as `input_source`. The GB300 SGLang tables used text-corpus inputs, so SGLang tables of the two campaigns are not corpus-comparable. Row labels are unchanged. - Each leaf's `fpm/provenance/provenance.json` lists the runs used (original, supplement, re-collected and superseded runs with reasons), the producer check against the source revision, platform notes, and every point not collected: beyond the measured KV capacity (`infeasible_kv_capacity`), a transient out-of-memory error at the deployment's single qualified memory setting although the KV fits (`infeasible_memory`), and excluded calibration and holdout points. Nothing is moved or imputed. - B300 vLLM FP8 TP4 is not part of this set: that profile is provided by [HF PR #19](https://huggingface.co/datasets/nvidia/aisimulate-fpm-dataset/discussions/19). - Not included: GB200 vLLM NVFP4 TP4 (holdout prefill MAPE 12.59%, above the gate). NVFP4 on Hopper, FP8 TP1 and TP8 on GB200 are design exclusions. Holdout FPM MAPE, prefill / decode (gate 10% per phase; all 38 phases pass): | System | Backend | Deployment | Prefill / decode | | --- | --- | --- | --- | | GB200 | vLLM 0.31.0 | FP8 TP4 | 6.09% / 1.84% | | GB200 | vLLM 0.31.0 | NVFP4 TP2 | 2.18% / 1.93% | | GB200 | SGLang 0.5.20 | FP8 TP4 | 1.55% / 1.65% | | GB200 | SGLang 0.5.20 | NVFP4 TP2 | 1.89% / 2.01% | | GB200 | SGLang 0.5.20 | NVFP4 TP4 | 2.13% / 1.16% | | B300 | vLLM 0.31.0 | FP8 TP2 | 1.55% / 2.40% | | B300 | vLLM 0.31.0 | NVFP4 TP1 | 1.48% / 2.01% | | B300 | vLLM 0.31.0 | NVFP4 TP2 | 1.45% / 1.51% | | B300 | vLLM 0.31.0 | NVFP4 TP4 | 1.67% / 1.00% | | B300 | SGLang 0.5.20 | FP8 TP2 | 1.66% / 2.27% | | B300 | SGLang 0.5.20 | FP8 TP4 | 0.87% / 1.51% | | B300 | SGLang 0.5.20 | NVFP4 TP2 | 1.20% / 2.05% | | B300 | SGLang 0.5.20 | NVFP4 TP4 | 1.74% / 1.43% | | H200 | vLLM 0.31.0 | FP8 TP4 | 1.23% / 1.38% | | H200 | vLLM 0.31.0 | FP8 TP8 | 1.34% / 1.13% | | H200 | SGLang 0.5.20 | FP8 TP4 | 0.79% / 2.13% | | H200 | SGLang 0.5.20 | FP8 TP8 | 2.88% / 1.15% | | H100 | vLLM 0.31.0 | FP8 TP8 | 3.00% / 1.53% | | H100 | SGLang 0.5.20 | FP8 TP8 | 1.85% / 1.51% |