AgentX FPM collection and dataset integration
Four Computelab FPM-on captures are integrated as measurement-only
configurations. They have standard flat configuration manifests, measurement
manifests, and catalog entries with fpm_file_count: 0. No performance library
is fabricated. Existing self-benchmark libraries and historical data are unchanged.
Configurations and data
| Experiment | Standard configuration leaf | Complete iterations | Incomplete groups |
|---|---|---|---|
| GLM-5.2 / B200 / SGLang, job 4208956 | pure-tp8 | 228,769 | 0 |
| DeepSeek-V4-Pro / B300 / SGLang, job 4209414 | dep8 | 98,227 | 0 |
| DeepSeek-V4-Pro / B200 / vLLM, job 4228930 | dep8 | 68,181 | 3,654 |
| MiniMax-M3 / B200 / vLLM, job 4245705 | pure-tp4 | 313,358 | 0 |
The complete lists contain 708,535 iterations and 1,873,391 rank observations. Another 23,759 rank observations remain as incomplete supporting evidence. All 1,897,150 retained FPM records are preserved, with no re-grouping or value changes during this layout migration.
Each leaf follows this layout:
<configuration>/
manifest.json # fpm: [], measurements reference
measurements/
manifest.json # protocol, representation and file hashes
fpm_iterations.json.gz # complete list[list[FPM]] truth
provenance/agentx-job-<job>/
incomplete_iterations.json.gz # supporting evidence, not scored
iteration_index.jsonl.gz # stored IDs/membership and array positions
alignment_report.json # run-specific grouping checks
collection_evidence.json # client config, validity and source times
processing-manifest.json # original processing inputs/counts/hashes
The processing manifest is the preserved original preparation artifact, not the
current configuration manifest. Its old relative filenames describe the source
package; the standard measurement manifest contains authoritative current paths.
Historical proposals/... strings occur only in source provenance referring to
the immutable preparation revision, not in active storage or discovery paths.
SGLang's installed package version was not retained. Its framework_version
therefore explicitly identifies the verified engine base and FPM patch commits:
git-71de97b264b0-fpm-ed18d64951b9. No release version is guessed. Full commits
and immutable image digests are in cases.json.
The vLLM cases retain their measured 0.28.0 version and distinct patch revisions.
aisim_commit is null because these are engine measurements, not an AISim
self-benchmark producer; the AISimulate reproduction-documentation commit is not
misrepresented as a producer runtime commit.
DeepSeek's weight_quantization: fp8_block records the pinned checkpoint's
quantization_config (fp8, dynamic activation, 128x128 weight blocks). This
label does not assert uniform tensor precision: the runtime uses FP4 routed
experts as documented by the collection. GLM and MiniMax use NVFP4 checkpoints.
Collection
All four runs used SemiAnalysisAI/aiperf
754356e9a39acc6cc6afb242d123bb57c3fb6f75: 393 Weka AgentX trajectories, seed42,
streaming, ignore_eos=true, first-turn-prefix cache busting, 10 warmup requests
per lane, and 3600 seconds of formal request sending. Sequence lengths follow
the replay, not a fixed ISL/OSL grid. Synthetic speculative acceptance measures
performance, not model quality.
- GLM: eight B200, attention TP8/DP1, EAGLE/native MTP, synthetic acceptance2.99, concurrency4, HiCache off; Dynamo frontend and SGLang worker.
- DeepSeek/B300: attention TP1/DP8, EP8, synthetic acceptance2.49, concurrency32, HiCache off; native SGLang and DP-aware consistent-hashing router.
- DeepSeek/B200: attention TP1/DP8, EP8, native vLLM MTP/V2, synthetic acceptance2.49, concurrency64, SimpleCPUOffloadConnector128GiB/rank; native Rust frontend.
- MiniMax/B200: four GPUs, TP4/DP1, EAGLE3-GQA/V1, synthetic acceptance2.78, concurrency15, G2 off, stream interval20; native Python frontend.
Native vLLM FPM is from PR52061, not Dynamo's InstrumentedScheduler. SGLang uses
the PR38711 compatibility backport. Each run has a separate buffered recorder,
with all DP endpoints observed by one recorder clock using time.time_ns().
Model revisions, image digests, backend/FPM SHAs, commands and reproduction links
are recorded in the case inputs and per-case provenance. Full collection
launchers remain in the immutable AISimulate experiment repository.
The original integration contained no prompts, generated text, model weights, full logs, raw request traces or credentials are added here. The dataset remains private. The migration preserves the source artifacts' provenance and does not grant new redistribution rights.
Trimming and grouping
The retained receipt-time interval is [profiling_start, profiling_start+3600s).
The exported end time includes drain and is not the cutoff.
| Job | Start, UTC | End exclusive, UTC |
|---|---|---|
| 4208956 | 2026-09-10 00:06:12.297086 | 2026-09-10 01:06:12.297086 |
| 4209414 | 2026-09-10 03:26:57.758555 | 2026-09-10 04:26:57.758555 |
| 4228930 | 2026-09-11 00:06:24.700077 | 2026-09-11 01:06:24.700077 |
| 4245705 | 2026-09-11 20:39:25.307076 | 2026-09-11 21:39:25.307076 |
GLM's naive export timestamp is PDT, verified against epoch-nanosecond request timestamps. The others are UTC. Filtering retains original integer nanoseconds, removes zero-time heartbeats, and keeps positive-time empty/dummy rank work. It excludes 26,951 pre-window records, 19,187 post-cutoff records and 1,993 in-window zero-time heartbeats from the original 1,945,281 records. Asynchronous publication makes boundary attribution approximate: this is not GPU-time slicing.
The JSON payload is a real outer iteration array with an inner rank array. DP1 uses singleton inner arrays. Native FPM values are unchanged; source line and receive timestamp are retained. Alignment was completed before publication:
- SGLang joins full-capture positive-time ordinals after equal-count, continuous-counter, global-correlation and 101-window-per-rank lag checks.
- vLLM uses monotone sequence alignment with duration/time anchors, two independent reference ranks and a conservative parameter-sensitivity check. Twenty-six ambiguous observations remain unresolved singletons. Missing dummy-step emissions are not fabricated.
There is no fixed 8ms membership rule. IDs are inferred, not native global execution IDs. Since duration is a vLLM alignment feature, post-alignment timing similarity is not independent proof of accuracy. The exact algorithm is preserved in alignment.py and the per-case alignment reports. The data migration itself only copies bytes.
Reading and validation
File entries declare representation: pre_grouped_rank_lists and file-level
grouping provenance under forward-pass-measurement-v1. Consumers read stored
membership; pre_grouped.py normalizes each inner list
to synchronizedIteration and uses maximum rank wall time without rematching.
import json, sys
from pathlib import Path
sys.path.insert(0, "scripts")
from pre_grouped import iter_canonical
root = Path(".")
config = json.loads(Path("data/deepseek-ai--DeepSeek-V4-Pro/b200-sxm/vllm/0.28.0/dep8/manifest.json").read_text())
bundle = json.loads((root / config["measurements"]["manifest_path"]).read_text())
truth = next(f for f in bundle["files"] if f["role"] == "truth")
iteration = next(iter_canonical(root, truth, list(range(config["dp"]))))
uv run --with pyarrow --with jsonschema python -m unittest discover -s scripts -p 'test_*.py'
uv run --with numpy==2.3.5 --with numba==0.65.0 python -m unittest discover -s scripts/agentx -p 'test_prepare.py'
uv run --with pyarrow --with jsonschema python scripts/manage_dataset.py validate
ingest_agentx.py migrates only the checksum-pinned
source package from HF revision 0d1bc03e053e5e327d2893943c5d2e99a993abed,
refuses configuration collisions and updates catalogs. It can be rerun on a
fresh baseline snapshot with the updated tooling; it intentionally refuses to
overwrite these already integrated leaves. The original preparation scripts are
kept for raw-source reproducibility, separate from canonical publication.
The base Playground import --replace now refuses to erase external campaigns;
rebuild in a fresh directory and reapply reviewed external ingestions.
See ingestion accounting and dataset validation. Dataset integration does not claim that an external FPM Gym deployment has already adopted this new representation; it provides the schema, normalization helper and discoverable data.
Request metric recovery
The manifest migration branch now includes the four capture-matched AIPerf request exports as metric-only compressed JSONL plus typed request Parquet. The original collection summary SHA-256 identifies each export; model and benchmark IDs, phase windows and completed request counts are checked again. The compressed records preserve original integer nanosecond timestamps and source order. No prompts or generated text are included.
See the request recovery audit for
all current and historical snapshots, including additional Playground captures.
Canonical measurement storage is now Parquet under measurement manifest v5;
measurement_storage.py dispatches its representation schema. Stored rank groups
and their order are unchanged. The earlier JSON description above documents the
original source package, which remains available at its pinned revision.