Add aggregated Inkling GB300 TP4, TP8 and DEP4 FPM data
Add the aggregated Inkling GB300 FPM calibration tables for TP4, TP8 and DEP4. This measurement-only PR includes completed, independently reviewed serving diagnostics for all three profiles and discloses the incomplete larger CPU replay. It does not claim broad serving-accuracy qualification.
The data covers thinkingmachines/Inkling-NVFP4 revision 42a75a99a40eb2ba1e0717db6357a0bf15205044, vLLM 0.27.0, NVFP4 weights and BF16 attention/KV. Each profile has three independent full-grid launches for prefill and three for decode. DEP8 is excluded.
| Profile | Retained prefill | Retained decode | Held coordinates |
|---|---|---|---|
| TP4 | 2,821 | 788 | 92 |
| TP8 | 2,740 | 938 | 180 |
| DEP4 | 2,429 | 548 | 132 |
The tables contain 10,264 retained coordinates. Each estimate is the median across three independent launches, after taking the maximum DP-rank latency within each launch. The original source measurement is excluded from that estimator. Retained coordinates pass the 5% CV gate; the 404 held coordinates remain in the portable proof populations. Three runs do not assess temporal split-half drift.
AISimulate #379 at 20e3eb9d52325a3e5a24e34c54dd09da130346fd loads these exact aggregate Parquet/metadata pairs without campaign files: all 10,264 stored latencies and 10,849 tested query results are unchanged. The original campaign publication/qualification revision 7d2c0a692ced4a433211819705958c04463dc146 remains recorded separately in version provenance.
Known limits remain explicit: the archived TP8 decode holdout has 35.3521% p95 relative error against a 20% target; TP4 and TP8 prefill each have one unsupported holdout. Loader compatibility and later coverage fixes do not replace these historical results. The measurement manifests intentionally contain no independent serving truth.
- Completed: independent repeated collection, robust aggregation, portable provenance, original publication validation and normal-loader compatibility.
- Attempted and incomplete: the 1,560-request CPU replay for TP4, TP8 and DEP4. Each exits with code 1 at its first observed unsupported prefill query: batch size 1, 1,536 prefill tokens, 253,952 KV-read tokens, without hitting a resource limit. Before stopping, TP4 resolves 1,541 measured and 427,477 interpolated queries; TP8 resolves 1,577 and 433,460; DEP4 resolves 2,049 and 473,890. These are native lookup resolutions; completed-request counts are unavailable from the failed CLI, and later coverage gaps remain unassessed. Nearby measurements at (prefill=1,024, KV=254,944) and (prefill=2,048, KV=253,920) do not provide the current direct interpolator's required axis bracketing. No extrapolation, interpolation change or new collection was used to obtain a pass.
- Completed: predictions and three matched-serving repetitions of one frozen 24-request serial AgentX play per profile (216 requests across the three profiles), with zero unsupported prediction queries and independent review of runtime, memory, protocol and result evidence. The ground truth is the pinned Inkling/vLLM runtime on GB300; AgentX supplies reconstructed workload inputs, not the original provider's timings.
| Profile | TTFT p95 relative error across repetitions | TPOT p95 relative error across repetitions |
|---|---|---|
| TP4 | 10.07–20.37% | 4.45–4.80% |
| TP8 | 12.21–16.05% | 4.46–4.52% |
| DEP4 | 20.87–21.02% | 8.67–10.09% |
All repetitions remain in the results. TP4's first TTFT result and all DEP4 TTFT results exceed the unchanged 20% reference threshold. DEP4 has observed attention-rank imbalance; no samples were excluded. The serial trace includes original idle gaps, so its throughput does not establish saturated serving capacity. HTTP/SSE timing boundaries and reconstructed prompts limit the comparison; no overhead correction was applied. TP8 ran on AWS GB300 hosts, while TP4 and DEP4 ran on OCI GB300 hosts, with a recorded driver difference and no claim of performance equivalence. These are campaign-local diagnostics, not canonical Stage 6 qualification, and they do not override the archived collection holdout results. The calibration metadata remains serving_accuracy: not_assessed; these diagnostics have not been added as independent truth files.
The PR adds the Inkling schema/import validation support and catalog entries together with the data. It is integrated with catalog v6: configuration v4, measurement v5, empty truth files and collection_runs: []. The public importer and validator check the reviewed portable import policy, snapshot/producer identity, provenance and catalog links. Worker precision, parallel configuration and CUDA graph evidence must agree with the published table identity; valid phase-specific graph behavior remains supported. The six large metadata/import JSON files use LFS with unchanged payload hashes. Existing dataset records and calibration values are preserved.
This PR is ready for formal review. The disclosed calibration, replay and serving-validation limits remain in effect. Consumer rollout requires AISimulate #379 or a version containing its loader support.
Validation progress: the first matched-serving attempt stopped during cold initialization, before any serving requests. Weight loading took about 573 seconds, followed by kernel compilation and autotuning. The 20-minute readiness deadline expired as CUDA graph capture began; no earlier fatal initialization exception was logged.
A reviewed bounded retry has been submitted with a 30-minute startup allowance and the original 45-minute total job cap. The frozen 24-request workload, model/runtime settings, CPU placement, memory admission checks and request limits are unchanged. DEP4 and TP8 remain gated on inspection of the TP4 smoke result.
Independent serving accuracy is still pending. The uploaded aggregate tables and disclosed replay/holdout limitations are unchanged.
Matched-serving validation update: TP4 completed all three repetitions of the frozen 24-request serial play (72 requests total). Independent review verified the model/runtime settings, physical cache layout and accepted memory budget, CPU affinity, exact request token IDs and output lengths, dispatch timing, completion records, and cache restoration after each play. No protocol or runtime-evidence defects were found.
| TP4 repetition | TTFT p95 absolute relative error | TPOT p95 absolute relative error |
|---|---|---|
| 1 | 20.3708% | 4.8020% |
| 2 | 10.1334% | 4.4481% |
| 3 | 10.0741% | 4.5084% |
Repetition 1 remains in the results and exceeds the unchanged 20% TTFT reference threshold. All three repetitions used one server initialization. Measured output throughput was 46.422–46.424 tokens/s versus 46.422 predicted, but this serial play includes source arrival times and idle gaps; that agreement does not establish saturated serving capacity. Client TTFT/TPOT also include the HTTP/SSE observation boundary; no overhead correction was applied.
The bounded DEP4 diagnostic has been submitted and is queued; TP8 follows after DEP4 completes and its results are inspected, with one GPU allocation at a time. These are campaign-local matched-serving diagnostics, not canonical Stage 6 qualification.
The uploaded aggregate tables are unchanged. The larger 12-play replay remains incomplete at the unsupported prefill query B=1, prefill tokens=1536, KV tokens=253952 for all three profiles. Previously disclosed collection holdout limitations, including the TP8 decode p95 error, remain in effect. This PR remains a draft.
The high-KV interpolation gap is accepted as a disclosed limitation for this publication. The larger 12-play replay remains incomplete at B=1, prefill tokens=1536, KV tokens=253952; later queries are unassessed. We are deferring interpolation changes and additional collection for that gap. This decision does not turn the replay into a passing validation or expand the serving-accuracy claim.
DEP4 completed three repetitions of the frozen 24-request matched-serving diagnostic (72 requests). Independent review verified the staged harness, runtime and precision/topology, grouped cache layout and accepted memory budget on all four ranks, CPU affinity, exact token IDs and output lengths, SSE completion, dispatch timing, scheduler linkage, and cache restoration. The latency metrics below were independently recomputed from the original evidence.
| DEP4 repetition | TTFT p95 absolute relative error | TPOT p95 absolute relative error |
|---|---|---|
| 1 | 20.8677% | 10.0877% |
| 2 | 21.0121% | 10.0109% |
| 3 | 21.0213% | 8.6738% |
All three TTFT p95 results exceed the unchanged 20% reference threshold and remain recorded. Observed per-rank scheduler activity is unequal: each active 100 ms bucket contains attention work from one DP rank, while each rank completes six requests per repetition. This is a serial-play result; cumulative request counts do not establish balanced simultaneous execution. The evidence does not measure expert routing or GPU overlap, and rank activity alone does not establish the cause of prediction error. Raw accuracy results are preserved.
The existing collection holdout limitations remain unchanged. All reported serving checks are campaign-local diagnostics, not canonical Stage 6 or saturation qualification. The dataset PR remains a draft.
The final TP8 diagnostic has been submitted with three repetitions and the reviewed one-hour wall limit. No other serving allocation is active. The uploaded measurement tables remain unchanged.
Reviewed PR head 76d4486b94e30ec2cfe6fafe4ea38c462c5c0d6a against main e62c2f64e158df608cb6701c02c4086cb4483d53 after PR #16 merged.
Findings to fix before merging
Update manifests and the importer to the merged schema. All three Inkling configurations fail the current schema checks: configuration manifests are v3 instead of v4; measurement manifests are v4 instead of v5 and omit
collection_runs; normalized configuration provenance still includes the now-rejectedsource_campaign_idandsource_campaign_path. Update the importer, regenerate the manifests and catalog v6 with dependent checksums, and resolve the catalog/README conflicts. These calibration-only bundles should retain empty truth files and usecollection_runs: [], without inventing evaluation runs. ImporterConnect Inkling admission checks to the new validation path. The hooks added in
manage_dataset.pyonly modify the legacy path. Catalog v6 delegates tomigrate_measurements.validate(), which currently sends every schema-7 FPM catalog record todsv41.validate_catalog_record(). I reproducedUnreviewed DSV41 source hash/revision/countfor TP4, TP8, and DEP4. Wire both Inkling snapshot/proof validation and catalog validation into the v6 path, dispatching by the applicable reviewed import policy and retaining rejection of unknown sources. Resolving merge conflicts alone will not fix this. Current validatorNormalize the six files declared as LFS but committed as regular Git objects. The three
fpm_forward_perf.metadata.jsonfiles and threefpm/provenance/import.jsonfiles match the new LFS rules but are stored as ordinary Git blobs.git lfs fsck --pointers HEADfails for all six; a fresh checkout continues to show them as modified after fetching the Inkling LFS assets. Re-add these six files through LFS without changing their payload bytes or content hashes. LFS rules
Validation and scope
- All three published evidence bundles passed the portable validator: TP4 3,609 rows, TP8 3,678 rows, DEP4 2,977 rows; 10,264 retained rows total.
- 74 synthetic tests passed. Two full-dataset cases were blocked by unhydrated legacy LFS payloads in the review checkout; the pinned-collector integration suite was skipped. Those environment limits are not reported as PR defects.
- Original GPU captures and independent serving accuracy were not revalidated. The disclosed replay gaps, holdout errors, and serving-validation limits remain in effect; passing the portable validator does not establish serving accuracy.
- Coordinate rollout with AISimulate #379, which was still open at review time and supplies the documented consumer compatibility.
Verdict: fix findings 1–3 before merging, then rerun dataset validation and importer tests against the merged v6 catalog. The calibration values and timing semantics should remain unchanged.
TP8 matched-serving validation is moving to AWS GB300 capacity. The original queued run was cancelled before GPU allocation. The replacement eight-GPU job is submitted on hold while the pinned checkpoint is staged; no TP8 serving result is available yet.
The model revision, vLLM 0.27.0 runtime image and overlays, frozen 24-request play, three repetitions, memory checks and accuracy reference thresholds are unchanged. The launcher independently passed review and will check the actual eight GPUs and their shared NVLink fabric before model startup. Driver, clocks and power are retained as evidence; performance equivalence between the two clusters is not assumed.
The uploaded FPM tables and prior TP4/DEP4 findings are unchanged. This remains a bounded serving diagnostic, with the previously disclosed high-KV replay gap and quality limitations retained. The dataset PR remains a draft.
TP8 matched-serving validation update (2026-10-13): the AWS GB300 run completed successfully in 29m55s, including initialization and all three repetitions of the frozen 24-request serial trace. The selected checkpoint, vLLM 0.27.0 runtime, precision, graph settings, workload, and comparison definitions were preserved. Hardware admission, all-eight-rank runtime/cache-capacity checks, and CPU-affinity checks passed.
| Repetition | Requests | TTFT p95 relative error | TPOT p95 relative error | Measured output tokens/s |
|---|---|---|---|---|
| 1 | 24 | 16.05% | 4.50% | 46.5566 |
| 2 | 24 | 12.21% | 4.46% | 46.5573 |
| 3 | 24 | 12.40% | 4.52% | 46.5600 |
These are p95 values of per-request absolute relative prediction error against serving measurements. All three TTFT and TPOT results are below the 20% reference threshold for this selected trace. Every repetition is retained. The predicted output rate was 46.5660 tokens/s; the measured 46.5566–46.5600 tokens/s includes the trace's original inter-turn idle gaps and is not a saturation-throughput result. Client HTTP/text-event measurements and simulated engine token timing have different boundaries; no HTTP-overhead subtraction was applied.
A nonfatal _seq_metadata_kernel JIT warning was observed during repetition 1. That repetition is retained; identical warm-up state across repetitions is not established.
The run used AWS GB300 hardware, while the FPM measurements were collected on OCI GB300 hardware. Hardware admission verified the intended GPU count and common NVLink fabric domain and retained driver/clock/power evidence; it does not establish performance equivalence between the environments.
This completes the planned small serving diagnostics for TP4, TP8, and DEP4. The earlier TP4 first-repetition TTFT miss and all three DEP4 TTFT misses remain recorded, together with the observed DEP4 scheduled-attention imbalance; no result was excluded. These diagnostics cover one serial play using reconstructed prompts and do not constitute canonical Stage 6 or broad serving-accuracy qualification. The larger 12-play replay remains incomplete at the disclosed high-KV interpolation gap, which was accepted for this publication scope; later queries remain unassessed. Existing collection-quality limitations, including TP8 decode holdout p95 error of 35.35% and the TP4/TP8 unsupported prefill holdouts, are unchanged.
The aggregate tables are unchanged. This dataset PR remains a draft.
Addressed findings 1–3 in this PR on top of main e62c2f64e158df608cb6701c02c4086cb4483d53: ada20b065490f1882f6d2d61609fa4bb33d5b3ef integrates catalog v6, b3951a217d781cf0125f297befa0ae77b21278bb normalizes LFS storage, and 5817f80c511848bb50e34df84c7cbd0ace059cf7 strengthens semantic validation.
- The importer writes configuration v4 and measurement v5, retaining empty truth files, a null protocol and
collection_runs: []. It generates catalog v6 directly without rerunning the historical measurement migration. All three bundles and their dependent checksums have been regenerated, and the catalog/README conflicts are resolved. Campaign identity remains checked in the portable evidence. - The v6 validator dispatches by reviewed import policy and verifies Inkling snapshots, portable proofs and catalog bindings. DSV41 and self-benchmark admission remain validated; unknown or altered policies, source identities, topology, provenance and catalog links are rejected.
- All six metadata/import JSON files are canonical LFS pointers. Their 53,890,836 payload bytes and SHA-256 values are unchanged. The six payloads were downloaded independently from Hugging Face into a new cache and their hashes verified. A fresh checkout is clean;
git lfs fsck --pointers HEADandgit lfs fsck --objects HEADpass, and all 751 tracked files match the reviewed checkout.
Independent whole-branch review also identified and closed runtime-evidence and normalized-identity consistency gaps. Precision/topology, graph evidence and repeated-launch settings must agree with the reviewed collection, and configuration/measurement/catalog snapshot and producer identities must match. Tests regenerate package hashes before checking semantic rejection, so hashes alone cannot hide those contradictions. Valid BF16 aliases and phase-specific graph behavior remain supported.
Validation on the final patch: the complete suite passed with 241 passed and 1 skipped. The skip is the optional pinned-collector exporter integration, whose collector module is unavailable in this dataset test environment. Both independent reviewers reran the fast suite (239 passed, 1 skipped, 2 long importer cases excluded) and approved the final branch; the complete run above includes those two importer cases. Full public dataset validation passes with 42 configuration snapshots and 291 measurement files. All 20 reviewer-authored invalid-package probes are rejected after package hashes are regenerated, and legitimate alias, phase-specific graph and observational-variation controls pass. The original separate 21-case admission-check matrix also passed.
All 27 Inkling timing/metadata/proof payloads, all 649 existing data files, the 59 existing canonical FPM records and all 291 measurement records are preserved. AISimulate #379 at 20e3eb9d52325a3e5a24e34c54dd09da130346fd returns the same 10,849 tested query results, including every one of the 10,264 published latencies.
Catalog regeneration recomputes the distinct source-file count from 216 to 256, matching the unchanged measurement records. It also restores optional DSV41/Kimi fields in FPM Parquet from main's existing canonical JSON; existing non-null values and underlying measurements are unchanged.
The README and opening description include the completed TP4, TP8 and DEP4 serving diagnostics. The disclosed holdout errors, accepted high-KV replay gap, driver difference and limits on serving-accuracy claims remain in effect. Consumer rollout requires AISimulate #379 or a version containing its loader support. This PR remains a draft for reviewer reassessment.
@nv-simonec , this PR is now ready for formal review at 5817f80c511848bb50e34df84c7cbd0ace059cf7. Could you please re-review the updated head? All three findings from your earlier review are addressed, and the timing data is unchanged.
The complete suite passes with 241 tests and one optional collector integration skipped. Independent reviews, full dataset validation, and fresh-checkout/LFS checks also pass. The opening description and prior reply contain the details.
The documented coverage and accuracy limitations remain in effect; consumer rollout is coordinated with AISimulate #379.