MANIFEST.md

Manifest — what was run

Without this file, nobody will be able to say six months from now under what conditions a given number was measured. It is the authoritative record of the runtime stacks, the harness settings and the uncommitted local changes in effect at measurement time.

This archive holds two batches. CANONICAL.md states the canonical run for every (version × model) cell.

Batch 1 · full clean sweep

run-id       vllm-{0.18.0,0.26.0,0.29.0}-vllm-only-h200-clean
date         2026-09-22 21:12 -- 2026-09-23 04:05 UTC (about 6h50m)
driver log   compare/fullsweep.log.gz
result       0.18 = 13/14, 0.26 = 13/14, 0.29 = 14/14
             (plus GLM-5.2-NVFP4 skipped on every leg, see below)

Batch 2 · gemma-4 on the 0.26 leg, transformers pinned

run-id       vllm-0.26.0-vllm-only-h200-tfpin   (1 model)
               PINS_026="transformers==5.14.1"
               venv: /data-fast/fk-venvs/venv-vllm-0.26.0-transformers-5.14.1
               scenario table patches/scenarios_gemma4.yaml
date         2026-09-23 05:33 UTC
driver log   compare/backfill.log.gz
merged       0.18 = 13/14, 0.26 = 14/14, 0.29 = 14/14

No cutlass backfill was required. On B200 a second batch pinned nvidia-cutlass-dsl==4.5.3 to recover six models on the 0.18 leg. On Hopper those six pass unpinned, so that batch does not exist here. This is itself a result — see "Failure attribution" in REPORT.md.

Hardware and node

node         alexsu-dev-h200-0
GPU          8 x NVIDIA H200, 143771 MiB each (1123 GiB total)
driver       NVRM 580.159.03
system CUDA  13.0 (V13.0.88)

For reference, the B200 node had 8 × 183359 MiB (1433 GiB). The tensor-parallel degrees were not changed — every model ran at the same tp as on B200, and nothing ran out of memory, including gpt-oss-120b at tp2, which was the candidate for trouble. So the two studies are directly comparable at the same tp.

nvidia/GLM-5.2-NVFP4 is skipped by the harness on this architecture (NVFP4 is Blackwell-only). It is a skip, not a failure: no process launched, so there is no run.log and no results.json for it on any leg.

The three runtime stacks

Identical to the B200 run, which is what makes the two comparable. torch / CUDA / flashinfer are printed by verify_one during setup and can be checked in compare/fullsweep.log.gz.

leg	vllm	torch	CUDA	flashinfer	transformers	cutlass-dsl
0.18	0.18.0	2.10.0+cu128	12.8	0.6.6	4.57.6	4.8.0
0.26	0.26.0	2.11.0+cu130	13.0	0.6.14	5.17.0	4.6.0
0.29	0.29.0	2.13.0+cu130	13.0	0.6.18	5.17.0	4.6.2

The pinned venv from batch 2 changes exactly one package:

0.26 + transformers 5.14.1   torch 2.11.0+cu130 / flashinfer 0.6.14

Two things worth recording. The 0.29 nvidia-cutlass-dsl version is 4.6.2; the B200 manifest lists it as "not recorded", so this fills that gap. And the 0.18 leg resolved nvidia-cutlass-dsl==4.8.0, exactly as on B200 — verified before the sweep that cute.core.ThrMma is absent from it. The broken dependency is present here too; it simply never gets imported on Hopper.

The three venvs resolve to different transformers and cutlass-dsl versions because uv pip install vllm==<ver> resolves each independently. This is a confound that can only be disclosed, not removed — 0.18 cannot use transformers 5.x.

Environment hygiene

• Three isolated uv venvs; the conda environment is only the driver. verify_one confirmed vllm-omni is absent from all three legs.
• All three legs have flashinfer-jit-cache and flashinfer-cubin at versions matching flashinfer-python. Without them, six MoE models fail on a gcc internal compiler error.
• VLLM_NO_USAGE_STATS=1 / DO_NOT_TRACK=1, and VLLM_LOGGING_LEVEL=DEBUG / PYTHONFAULTHANDLER deliberately not set: DEBUG output corrupts the stdout the usage-reporting thread parses, and ray_runner._is_fatal_worker_line() then reads the resulting traceback as a crash and kills a healthy worker.
• Weights were fetched by patches/prefetch_weights.sh before the sweep (1.6 TB, 15 models, zero incomplete files), so download time is not inside any task's elapsed_s.

One deviation, the same as on B200: UV_CACHE_DIR resolved to /tmp/cursor-sandbox-cache/.../uv rather than /data-fast/uv-cache. This only affects venv install speed (uv cannot hardlink, so wheels are copied in full) and does not affect measurements.

Harness settings

scenario table  fastkernels/scenarios/vllm_only.yaml, 15 models
                (archived as patches/vllm_only.yaml; 14 run, GLM skipped)
execution       strictly serial, whole node per task (FASTKERNELS_VALIDATE_SERIAL=1)
                no --resume, so every number is a real measurement
results dir     /code/users/alexsu/vllm-validate-h200   (deliberately separate
                from the B200 tree so the two studies cannot mix)
repository      main @ 9acebaa1349c78af8bb0ecb7e0719b5372d26fd8
                plus the same three uncommitted local changes as the B200 run

Local changes (uncommitted)

The same three as the B200 study; the patches are in patches/ and are the only reliable record. Verified present on this node before the sweep:

grep -c 'getattr(mnnvl, "IpcSocket", None)' fastkernels/validate/bench_vllm.py   # 4
grep -c '_WHISPER_WINDOW_S'                 fastkernels/validate/bench_vllm.py   # 3
grep -c FASTKERNELS_VALIDATE_SERIAL         fastkernels/validate/ray_runner.py   # 1

1. patches/bench_vllm.py.patch — the IpcSocket fix (four call sites) and the whisper 30s clamp. LibriSpeech test.clean does contain samples longer than 30s; the clamp works around a vLLM 0.29 whisper bug and must be disclosed.
2. patches/ray_runner.py.patch — the FASTKERNELS_VALIDATE_SERIAL switch.
3. patches/compare_vllm_versions_datafast.sh and patches/prefetch_weights.sh — untracked driver scripts, archived in full.

Archive contents

README.md                          the dataset card, generated from REPORT.md
REPORT.md                          analysis report
MANIFEST.md                        this file
CANONICAL.md                       canonical run per (version x model)
make_tables.py                     regenerates the report tables from results/
make_card.py                       regenerates README.md from REPORT.md

results/<run-id>/<task>/results.json   41 files
results/<run-id>/run.jsonl             event stream per run

compare/0.18-vs-0.26.txt           pairwise tables, generated from the canonical
compare/0.26-vs-0.29.txt             run set (same source as REPORT.md's tables)
compare/fullsweep.log.gz           batch 1 driver log
compare/backfill.log.gz            batch 2 driver log

logs.tar.gz                        43 run.log (14 models x 3 legs + 1 tfpin;
                                   GLM was skipped so it produced none)
patches/                           uncommitted local changes + scenario tables

Re-checking the pass matrix

python3 - <<'PY'
import json
p = 'results/vllm-0.29.0-vllm-only-h200-clean/run.jsonl'
subs, fins = {}, []
for line in open(p):
    r = json.loads(line)
    if r.get('event') == 'task_submitted':
        subs[r['job']['index']] = r['job']['name']
    if r.get('event') == 'run_finished':
        fins.append(r['results'])
for i, s in sorted(fins[-1].items(), key=lambda kv: int(kv[0])):
    print(f'{i:>3} {s:22} {subs.get(int(i))}')
PY

SKIP(nvfp4-hopper) on index 2 is GLM-5.2-NVFP4 and is expected on every leg.