Alexsssu/fastkernels-validate-h200-20260923

FastKernels full validate on 8×H200 — 2026-09-23

End-to-end results of a complete FastKernels validate full sweep on 8×NVIDIA H200 SXM 141GB: all 47 scenarios of fastkernels/scenarios/full.yaml across 22 benchmark harnesses, comparing FastKernels against each architecture's production or upstream reference.

46 / 46 submitted scenarios produced results. Across the 45 architectures with a throughput ratio, FastKernels reaches a geometric mean of 1.205× against the reference implementations, median 1.043×, with 36/45 at or above parity.

Results

Category	Architecture	Reference	Dataset / Workload	Throughput	Latency	Align.
Dense & MoE LLMs	meta-llama/Llama-3.1-8B-Instruct	vLLM	wildchat-mixed-1k + longbench-longctx mixed, long-context	0.974x	1.004x	exact 274/1000 (27.4%)
	mistralai/Mixtral-8x7B-Instruct-v0.1	vLLM	wildchat-mixed-1k + longbench-longctx mixed, long-context	1.020x	1.016x	exact 131/1000 (13.1%)
	microsoft/bitnet-b1.58-2B-4T	microsoft-bitnet-gpu	wildchat-mixed-1k + longbench-longctx mixed, long-context	1.015x	1.097x	exact 5/32 (15.6%)
	openai/gpt-oss-120b	vLLM	wildchat-mixed-1k + longbench-longctx mixed, long-context	1.038x	0.923x	exact 146/1000 (14.6%)
	meta-llama/Llama-3.1-8B-Instruct †	SGLang	— eagle3-16seqs-out256	0.965x	1.044x	exact 8/16 (50.0%)
	google/gemma-4-26B-A4B-it	vLLM	wildchat-mixed-1k + longbench-longctx mixed, long-context	0.921x	1.030x	exact 146/1000 (14.6%)
Linear Attention & New Archs	state-spaces/mamba-2.8b-hf	vLLM	wildchat-mixed-1k + longbench-longctx mixed, long-context	1.085x	1.026x	exact 379/1000 (37.9%)
	mistralai/Mamba-Codestral-7B-v0.1	vLLM	wildchat-mixed-1k + longbench-longctx mixed, long-context	0.996x	0.962x	exact 444/1000 (44.4%)
	fla-hub/rwkv7-2.9B-g1	FLA	wildchat-mixed-1k + longbench-longctx mixed, long-context	1.350x	1.297x	exact 176/1000 (17.6%)
	fla-hub/gla-2.7B-100B	FLA	wildchat-mixed-1k + longbench-longctx mixed, long-context	1.348x	1.670x	exact 702/1000 (70.2%)
	fla-hub/retnet-2.7B-100B	FLA	wildchat-mixed-1k + longbench-longctx mixed, long-context	3.632x	1.757x	exact 51/1000 (5.1%)
	Qwen/Qwen3-Next-80B-A3B-Instruct	vLLM	wildchat-mixed-1k + longbench-longctx mixed, long-context	1.022x	0.992x	exact 26/1000 (2.6%)
	moonshotai/Kimi-Linear-48B-A3B-Instruct	vLLM	wildchat-mixed-1k + longbench-longctx mixed, long-context	1.078x	0.892x	exact 79/1000 (7.9%)
	ttt_e2e	JAX ref.	— ttt-e2e-pretrain, ttt-e2e-meta	1.962x	1.962x ‡	token_nll_mean_abs_diff 0.0226
	ai21labs/AI21-Jamba-Mini-1.7 †	vLLM	wildchat-mixed-1k + longbench-longctx mixed, long-context	0.961x	1.065x	exact 155/1000 (15.5%)
Vision / Video / Audio	black-forest-labs/FLUX.1-dev	vllm-omni	nateraw/parti-prompts 1024x1024, 512x512	1.114x	—	min_cosine_sim 0.9813
	hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v	vllm-omni	movie_gen_video_bench 480p-short, 480p-medium	1.001x	—	min_cosine_sim 0.9740
	stabilityai/stable-diffusion-xl-base-1.0	diffusers	Parti-prompts 1024x1024, 512x512	—	—	min_cosine_sim 0.9839
	facebook/sam3.1	facebook/sam3	SACo-VEval smartglasses_val full-pipeline, smartglasses-val-video	1.130x	1.094x	min_cosine 0.8713
	openai/whisper-large-v3	vLLM	librispeech_asr test.clean librispeech	1.041x	0.950x	exact 1667/2620 (63.6%)
	FunAudioLLM/Fun-CosyVoice3-0.5B-2512	vllm-omni	SEED-TTS-Eval tts-short, tts-medium, tts-long	2.853x	—	min_mel_cosine_sim 0.7433
Multimodal & Encoders	Qwen/Qwen2-VL-7B-Instruct	vLLM	— text-only, image, video	1.137x	0.989x	exact 361/1000 (36.1%)
	Qwen/Qwen3-VL-8B-Instruct	vLLM	— text-only, image, video	1.227x	1.020x	exact 195/1000 (19.5%)
	Qwen/Qwen3-VL-235B-A22B-Instruct-FP8	vLLM	— text-only, image, video	1.562x	0.978x	exact 78/1000 (7.8%)
	Qwen/Qwen2.5-Omni-7B	vLLM	— text, image, video, audio	1.517x	1.014x	exact 314/1000 (31.4%)
	google/siglip2-so400m-patch16-naflex †	timm	ILSVRC/imagenet-1k default-res, high-res	1.000x	—	min_cosine_sim 0.9923
	facebook/dinov3-vit7b16-pretrain-lvd1689m †	timm	ILSVRC/imagenet-1k default-res, high-res	0.948x	—	min_cosine_sim 1.0000
	timm/swinv2_large_window12_192.ms_in22k †	timm	ILSVRC/imagenet-1k default-res, high-res	1.005x	—	min_cosine_sim 1.0000
Edge & Detection	timm/mobilenetv4_conv_medium.e500_r256_in1k †	timm	ILSVRC/imagenet-1k default-res, high-res	1.044x	—	min_cosine_sim 1.0000
	facebook/convnextv2-base-22k-384	transformers	ethz/food101 imagecls-ethz/food101	1.001x	0.988x	top1_match_rate 1.0000
	timm/efficientnetv2_rw_m.agc_in1k	timm	ethz/food101 imagecls-ethz/food101	1.061x	0.993x	top1_match_rate 1.0000
	jameslahm/yolov10n	yolov10	COCO val2017 coco-val	1.043x	—	boxes_cosine 0.4611 ⚠
	PekingU/rtdetr_v2_r101vd	transformers	COCO val2017 coco-val	1.001x	—	boxes_cosine 0.9279 ⚠
3D / Robotics / Science	Voxel51/gaussian_splatting	gsplat	100k-Gaussian orbit render render	1.005x	0.999x	rgb_cosine 1.0000
	NVlabs/instant-ngp	pyngp	fox render benchmark render	0.989x	0.989x	rgba_cosine 1.0000
	Pointcept/PointTransformerV3	PTv3 ref.	scanobjectnn nobg_test scanobjectnn, batch-8	1.001x	0.991x	feat_cosine 0.9862 ⚠
	OpenFold/OpenFold3	OpenFold3 ref.	OpenProteinSet short, medium, long, extra-long	1.002x	0.957x	pass_rate 1.0000
	pi0_aloha_pen_uncap	OpenPI	ALOHA real demos aloha-3cam, aloha-1cam	3.006x	3.739x	mean_cosine_sim 1.0000
	dp3	3D-Diffusion-Policy	gym-xarm point clouds dp3-1env, dp3-batch	1.138x	—	mean_cos 1.0000
Recommendation & Specialized	dlrmv2	TorchRec DLRM	Adult Census ctr-batch	1.072x	1.337x	min_cosine 0.0000 ⚠
	lightgcn	PyG LightGCN	MovieLens recommend-batch	1.023x	1.037x	min_cosine 1.0000
	BAAI/bge-m3	vLLM token_embed	sentence-transformers/mldr bge-m3-mldr-docs	1.375x	1.042x	min_cosine 0.9971
	colbert-ir/colbertv2.0	vLLM token_embed	MS MARCO passages colbertv2-msmarco-passages	3.771x	2.649x	min_cosine 1.0000
	GSAI-ML/LLaDA-8B-Instruct	Fast-dLLM	HumanEval humaneval-fastdllm-dual	1.044x	1.046x	exact 88/164 (53.7%)
World Models	Etched/oasis-500m	open-oasis	Minecraft-VLA clips 4 ddim configs	1.180x	1.132x	min_cosine 0.9983
	facebook/vjepa2-vitl-fpc64-256	transformers	nateraw/kinetics-mini predictor, encoder	0.994x	1.002x	min_cosine 1.0000

Throughput vs reference across 45 architectures — geometric mean 1.205x, median 1.043x, range 0.921x – 3.771x. At or above parity: 36/45 · within 5% of parity: 23/45 · above 1.20x: 11/45. Excluding the 4 scenarios whose correctness check is unreliable (⚠), the geometric mean over the remaining 41 is 1.224x.

Legend — ⚠ the harness-internal correctness check reports a failure; all four are harness bugs, not kernel regressions (see below). † produced by a serial rerun on a different machine, so not in the same measurement condition as the other rows. ‡ this harness emits identical figures for throughput and latency, so the two columns are not independent. A — in Dataset means the harness does not record the corpus in results.json.

Where the wins and losses are

Strongest — Pi0 robotics policy 3.01×, ColBERTv2 late-interaction retrieval 3.77×, RetNet 3.63×, CosyVoice3 TTS 2.85×, TTT-E2E 1.96×, Qwen3-VL-235B-FP8 1.56×, Qwen2.5-Omni 1.52×. The pattern is consistent: the largest gains are on under-served architectures where the upstream reference is a research codebase rather than a hardened serving engine.

At parity — mainstream LLM serving against vLLM and SGLang lands within a few percent: Llama-3.1 0.974×, Mixtral 1.020×, gpt-oss-120b 1.038×, Mamba-Codestral 0.996×, EAGLE-3 vs SGLang 0.965×. 23 of 45 architectures sit within ±5% of parity.

Below parity — Gemma-4 0.921× is the largest shortfall, followed by DINOv3 0.948×, Jamba 0.961×, V-JEPA 2 0.994×, InstantNGP 0.989×.

Correctness

Alignment against the reference is clean for the large majority: exact bit-identical output (MSE 0.0) on DINOv3, SwinV2 and MobileNetV4, cosine ≥ 0.999 on ColBERTv2, LightGCN, V-JEPA 2, OpenFold3 (pass_rate 1.0), Pi0 and DP3, and ≥ 0.98 on FLUX, SDXL, HunyuanVideo and Oasis.

Four scenarios report an internal correctness failure. All four are harness defects, not kernel regressions, and each has a different cause:

Scenario	Reported	Actual cause
jameslahm/yolov10n	boxes_cosine 0.461	The reference adapter reads the raw [B, 84, 8400] detection head as if it were the end2end [x1,y1,x2,y2,conf,cls] export. The reference "scores" reach 118.4 and "labels" reach 127 (COCO has 80 classes) — the comparison is against garbage.
PekingU/rtdetr_v2_r101vd	labels_match_rate 0.734	Both sides are valid. The metric is unweighted across all 100 slots, ~95 of which are sub-0.01-confidence noise. Restricted to real detections: top-1 label match 99.73%, median top-1 box difference 0.0000.
Pointcept/PointTransformerV3	feat_cosine 0.9862 vs 0.99	Threshold mis-set: the FastKernels paper's own published value for this architecture is 0.9796, which this threshold would also fail.
dlrmv2	min_cosine 0.0	Self-contradictory — reported alongside max_abs_diff 0.0, i.e. bit-identical outputs. Caused by the zero-norm guard in _safe_cosine_similarity.

A related finding: status=PASS in the run log only means exit code 0. The runner sets status = "PASS" if returncode == 0; a harness's internal overall_pass / alignment.passed does not participate. That is why these four appeared green in the run summary.

Failures and how they were resolved

The original run reported 40 PASS / 6 FAIL / 1 SKIP. All 6 failures were investigated and resolved; none was a kernel defect.

Scenario	Cause	Resolution
4 × bench_timm (SigLIP-2, DINOv3, SwinV2, MobileNetV4)	ILSVRC/imagenet-1k gated dataset access not granted	Access granted, all four rerun clean
bench_sglang EAGLE-3	SGLang 0.5.20 turned ServerArgs into a msgspec Struct; dataclasses.fields() raises. Reproduced verbatim on a freshly built env — a real incompatibility, not a broken install	Two-line patch (scripts/bench_sglang.msgspec.patch)
AI21-Jamba-Mini-1.7	Spurious CUDA error: an illegal memory access under parallel scheduling, crashing on the first inference step (0/1000) right after CUDA graph capture	Passed on a serial rerun with no code change

The Jamba result means the parallel execution mode produces spurious failures. Since parallel is the mode the driving command specifies, the throughput figures from that mode may also carry concurrency-induced dips; this run did not quantify that.

nvidia/GLM-5.2-NVFP4 is the single SKIP — NVFP4 is Blackwell-only, so skipping on Hopper is correct behaviour, not a failure.

Environment

GPU	8 × NVIDIA H200 SXM 141GB
Driver / CUDA	580.159.03 / 13.0
GPU clocks	not locked
torch	2.11.0+cu130
vllm	0.26.0
vllm-omni	0.26.0rc1
flash-attn	2.8.3
transformers	5.14.1
sglang (isolated env)	0.5.20 / torch 2.13.0+cu130
Repo	main @ 9acebaa, pristine clone, no local changes for the original run
Wall time	108.5 min for 46 parallel jobs

The driving command was executed verbatim — no --output-dir, no timeout overrides, no serial override — so Ray scheduled all 46 jobs in parallel as intended.

Relation to the FastKernels paper

These numbers are not a reproduction of arXiv 2605.23215 and should not be read as one. The paper measures on H100 SXM5 80GB with clocks locked at 1,593 MHz; this run is unlocked H200. Across 44 comparable architectures no value matches exactly, the median deviation is 7.6%, and 9 architectures flip direction (one side above parity, the other below).

The paper anticipates exactly this: "rankings on B200, MI300X, or consumer GPUs may differ and should be re-run on the target hardware." The largest deviations land on bandwidth-bound architectures (RetNet +95%, TTT-E2E +78%, CosyVoice3 +34%), consistent with H200's ~43% higher HBM bandwidth. Protocol also differs: the paper averages 3 runs with 10 warmup iterations and uses synthetic tensors for detection, whereas this run times once and uses real COCO images.

Full side-by-side in tables/paper_comparison.md; discussion in the reports.

Reports and data

Full analysis: REPORT.en.md (English) · REPORT.md (中文). Run conditions, hardware, software stack, local changes and deviations: MANIFEST.txt.

tables/
  summary.md                    The table above, regenerable
  coverage.md                   Per-scenario status, data source, internal verdict
  throughput.md                 Per-scenario throughput ratio
  latency.md                    Per-scenario latency speedup
  alignment.md                  All alignment / similarity metrics
  paper_comparison.md           Side-by-side against the paper's Tables 7/8
results/
  orig/run.jsonl                Event stream of the original parallel run
  orig/<NNN_harness_model>/results.json     40 scenarios
  rerun/<NNN_harness_model>/results.json     6 scenarios (serial reruns)
scripts/
  run.sh                        Driver for the original run (command as given)
  run_reruns.sh                 Driver for the serial reruns
  prefetch_full.sh              Weight prefetch for the full scenario table
  prefetch_rerun.sh             Weight prefetch for the rerun subset
  setup_node.sh                 Node environment setup
  env.sh                        Environment variables
  make_tables.py                Regenerates tables/ from results.json
  bench_sglang.msgspec.patch    The one local change, needed for the SGLang scenario
logs.tar.gz                     All run.log files plus console output (58 files)

Intermediate comparison artifacts (feats/, *_raw.json, generated images / videos / point clouds) are not included — 26 GB in the source tree, 20 GB of it from SAM 3.1 alone. They are scratch data for correctness comparison, not results; rerun the relevant scenario to regenerate.

Caveats

• Every figure is a single timing. No repeat measurement, no error bars. Differences of ±5% should not be read as real.
• GPU clocks were not locked — the largest methodological gap in this run.
• 15 bench_vllm scenarios carry the vllm-omni plugin, which pyproject.toml declares a hard dependency and which injects into every vLLM process via [vllm.general_plugins]. Three scenarios genuinely require it, so this is expected for this scenario table — but those 15 vLLM numbers are not comparable against clean, omni-free measurements.
• The 6 rerun rows (†) came from a different physical machine under serial scheduling and are not in the same measurement condition as the rest.
• The 4 harness bugs are diagnosed but not fixed. Correctness verdicts for those scenarios are unusable until they are.

Citation

@misc{oliaro2026fastkernels,
      title={FastKernels: Benchmarking GPU Kernel Generation in Production},
      author={Gabriele Oliaro and Yichao Fu and May Jiang and Owen Lu and Junli Wang and Zhihao Jia and Hao Zhang and Samyam Rajbhandari},
      year={2026},
      eprint={2605.23215},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.23215},
}