Alexsssu/fastkernels-h200-vllm-ablation

vLLM 0.18 → 0.26 → 0.29 on H200: ablation, a three-way comparison

2026-09-22, 8×H200 (alexsu-dev-h200-0). How it was run, the runtime stacks and the uncommitted local changes are all in MANIFEST.md.

This is the Hopper companion to the B200 ablation of 2026-09-21. The scenario table, the harness, the tensor-parallel degrees and all three runtime stacks are byte-for-byte the same as the B200 run, so the two are directly comparable. The headline result is that most of the B200 conclusions do not survive the change of architecture; see "What transfers from B200, and what does not".

Throughput is a single timed pass per scenario, with no repeats. See "Known limitations".

About this dataset

This page is the report. It is generated from REPORT.md; the two are always identical.

Companion study: the same ablation on B200 (Blackwell) is at Alexsssu/fastkernels-b200-vllm-ablation. Same scenario table, same tensor-parallel degrees, same three runtime stacks -- only the GPU differs, which is what makes the two comparable. Most of the B200 conclusions do not reproduce here; see "What transfers from B200, and what does not" below.

path	what it is
REPORT.md	the analysis below, as a file
CANONICAL.md	which run is authoritative for each (version x model), and what was excluded
MANIFEST.md	runtime stacks, harness settings, uncommitted local changes
results/	41 results.json plus the per-run run.jsonl event streams
compare/	pairwise comparison tables, built from the same canonical runs as the tables below
logs.tar.gz	43 run.log, one per task
patches/	the uncommitted local changes, the scenario tables, the driver scripts
raw/	raw harness output (114 MB tarball); see raw/RAW_CONTENTS.md
make_tables.py, make_card.py	regenerate the tables from results/, and this page from REPORT.md

Every number below comes from results/. If the tables, compare/*.txt and results/ ever disagree, that is a bug -- please open a discussion.

Coverage

leg	passing	not run
0.18.0	13/14	gemma-4-26B
0.26.0	14/14	—
0.29.0	14/14	—

The denominator is 14, not 15. nvidia/GLM-5.2-NVFP4 is skipped on every leg, not failed: the harness refuses it on Hopper with NVFP4 is Blackwell-only. It therefore has no row anywhere in this report.

13 models support a three-way comparison. gemma-4-26B is two-way (0.26 and 0.29 only).

Headline findings

• vLLM 0.18 is not broadly handicapped on Hopper. It reaches 13/14 with no dependency pins at all. On B200 the same version reached only 7/15 in the clean sweep and needed a nvidia-cutlass-dsl pin to get to 13/15. All six models that the pin rescued on B200 pass here untouched.
• The 0.26 regression on Qwen3-VL-235B-FP8 reproduces almost exactly. Image throughput 6,539 → 4,614, i.e. 0.71x, against 0.70x on B200. This is the only B200 finding that transfers cleanly, and the agreement is the strongest evidence yet that it is a real vLLM 0.26 defect rather than a kernel or hardware artifact.
• The big 0.29 wins are Blackwell-only. Qwen3-Next-80B 1.76x → 1.01x, Kimi-Linear-48B 1.69x → 1.13x, Qwen3-VL-235B-FP8 image 2.08x → 1.17x. On Hopper, 0.26 → 0.29 is close to flat everywhere.
• gpt-oss-120b's 3.31x disappears. On B200 this was the single largest 0.18 → 0.26 jump and was read as mxfp4 support maturing. On H200 it is 1.03x — because 0.18 is already fast here (11,290 tok/s vs 7,304 on B200). The jump measured a Blackwell-specific deficiency in 0.18, not a general improvement in 0.26.
• 0.18 has a long-context deficit that only appears on Hopper. Llama-3.1-8B 1.53x and AI21-Jamba-Mini-1.7 1.78x on 0.18 → 0.26, where B200 showed 1.00x and 1.14x. See "The 0.18 long-context deficit".

Full data

Throughput (tok/s, higher is better)

model	tp	workload	0.18	0.26	0.29	26/18	29/26
Llama-3.1-8B-Instruct	1	mixed	12,779	12,931	12,849	1.01x	0.99x
Llama-3.1-8B-Instruct	1	long-context	109	167	167	1.53x	1.00x
Mixtral-8x7B-Instruct-v0.1	2	mixed	10,182	10,147	10,171	1.00x	1.00x
Mixtral-8x7B-Instruct-v0.1	2	long-context	343	344	342	1.00x	0.99x
gpt-oss-120b	2	mixed	11,290	11,635	11,698	1.03x	1.01x
gpt-oss-120b	2	long-context	370	390	378	1.05x	0.97x
mamba-2.8b-hf	1	mixed	7,463	8,718	8,669	1.17x	0.99x
mamba-2.8b-hf	1	long-context	255	266	266	1.04x	1.00x
Mamba-Codestral-7B-v0.1	1	mixed	4,368	7,209	7,649	1.65x	1.06x
Mamba-Codestral-7B-v0.1	1	long-context	286	291	294	1.02x	1.01x
Qwen3-Next-80B-A3B-Instruct	2	mixed	9,205	9,096	9,200	0.99x	1.01x
Qwen3-Next-80B-A3B-Instruct	2	long-context	433	456	464	1.05x	1.02x
AI21-Jamba-Mini-1.7	1	mixed	3,416	3,883	3,896	1.14x	1.00x
AI21-Jamba-Mini-1.7	1	long-context	65	117	117	1.78x	1.00x
Kimi-Linear-48B-A3B-Instruct	2	mixed	11,582	11,515	13,014	0.99x	1.13x
Kimi-Linear-48B-A3B-Instruct	2	long-context	609	598	643	0.98x	1.07x
whisper-large-v3	1	librispeech	7,732	7,910	7,889	1.02x	1.00x
Qwen2-VL-7B-Instruct	1	text-only	11,241	11,188	11,244	1.00x	1.01x
Qwen2-VL-7B-Instruct	1	image	9,211	9,766	10,147	1.06x	1.04x
Qwen2-VL-7B-Instruct	1	video	1,612	1,674	1,905	1.04x	1.14x
Qwen3-VL-8B-Instruct	1	text-only	8,773	8,786	8,864	1.00x	1.01x
Qwen3-VL-8B-Instruct	1	image	8,005	8,255	8,225	1.03x	1.00x
Qwen3-VL-8B-Instruct	1	video	1,653	1,687	1,716	1.02x	1.02x
Qwen3-VL-235B-A22B-Instruct-FP8	4	text-only	3,717	3,101	3,263	0.83x	1.05x
Qwen3-VL-235B-A22B-Instruct-FP8	4	image	6,539	4,614	5,400	0.71x	1.17x
Qwen3-VL-235B-A22B-Instruct-FP8	4	video	1,148	1,128	1,165	0.98x	1.03x
Qwen2.5-Omni-7B	1	text	11,229	11,248	11,270	1.00x	1.00x
Qwen2.5-Omni-7B	1	image	9,140	9,251	9,334	1.01x	1.01x
Qwen2.5-Omni-7B	1	video	1,500	1,505	1,538	1.00x	1.02x
Qwen2.5-Omni-7B	1	audio	3,678	4,559	5,022	1.24x	1.10x
gemma-4-26B-A4B-it	1	mixed	—	8,471	8,955	—	1.06x
gemma-4-26B-A4B-it	1	long-context	—	201	208	—	1.03x

Latency (ms/tok, lower is better)

model	tp	workload	0.18	0.26	0.29	18→26	26→29
Llama-3.1-8B-Instruct	1	single-request	5.10	5.02	5.02	1.01x	1.00x
Llama-3.1-8B-Instruct	1	fixed-batch-32	0.24	0.24	0.24	1.00x	1.00x
Mixtral-8x7B-Instruct-v0.1	2	single-request	5.39	5.35	5.34	1.01x	1.00x
Mixtral-8x7B-Instruct-v0.1	2	fixed-batch-32	0.50	0.51	0.50	0.99x	1.01x
gpt-oss-120b	2	single-request	5.30	3.97	4.12	1.33x	0.96x
gpt-oss-120b	2	fixed-batch-32	0.36	0.32	0.33	1.14x	0.96x
mamba-2.8b-hf	1	single-request	3.75	3.71	3.71	1.01x	1.00x
mamba-2.8b-hf	1	fixed-batch-32	0.41	0.41	0.41	1.00x	1.00x
Mamba-Codestral-7B-v0.1	1	single-request	5.50	5.46	5.36	1.01x	1.02x
Mamba-Codestral-7B-v0.1	1	fixed-batch-32	0.42	0.32	0.32	1.32x	1.01x
Qwen3-Next-80B-A3B-Instruct	2	single-request	5.55	6.47	5.07	0.86x	1.28x
Qwen3-Next-80B-A3B-Instruct	2	fixed-batch-32	0.42	0.40	0.38	1.03x	1.06x
AI21-Jamba-Mini-1.7	1	single-request	8.96	8.79	8.78	1.02x	1.00x
AI21-Jamba-Mini-1.7	1	fixed-batch-32	1.14	1.14	1.14	1.00x	1.00x
Kimi-Linear-48B-A3B-Instruct	2	single-request	4.41	4.06	3.29	1.08x	1.24x
Kimi-Linear-48B-A3B-Instruct	2	fixed-batch-32	0.36	0.35	0.32	1.03x	1.08x
whisper-large-v3	1	single-utterance	2.64	2.52	2.53	1.05x	1.00x
whisper-large-v3	1	fixed-batch-32	0.18	0.18	0.18	1.00x	1.00x
Qwen2-VL-7B-Instruct	1	single-image	5.13	5.21	5.14	0.98x	1.01x
Qwen2-VL-7B-Instruct	1	single-video	7.05	7.06	6.58	1.00x	1.07x
Qwen3-VL-8B-Instruct	1	single-image	5.55	5.66	5.53	0.98x	1.02x
Qwen3-VL-8B-Instruct	1	single-video	7.16	7.22	7.01	0.99x	1.03x
Qwen3-VL-235B-A22B-Instruct-FP8	4	single-image	12.22	9.64	9.19	1.27x	1.05x
Qwen3-VL-235B-A22B-Instruct-FP8	4	single-video	14.90	12.18	11.27	1.22x	1.08x
Qwen2.5-Omni-7B	1	single-text	4.95	4.98	4.95	0.99x	1.01x
Qwen2.5-Omni-7B	1	single-image	5.23	5.26	5.23	0.99x	1.01x
Qwen2.5-Omni-7B	1	single-video	7.26	7.34	7.17	0.99x	1.02x
Qwen2.5-Omni-7B	1	single-audio	5.20	5.36	5.29	0.97x	1.01x
gemma-4-26B-A4B-it	1	single-request	—	4.46	4.39	—	1.02x
gemma-4-26B-A4B-it	1	fixed-batch-32	—	0.44	0.42	—	1.04x

What transfers from B200, and what does not

Same scenario table, same tp, same three runtime stacks, same harness, same serial-exclusive execution. The only variable is the GPU. Speedups that differ between the two architectures by more than 1.25x:

model / workload	comparison	B200	H200
gpt-oss-120b mixed	26/18	3.31x	1.03x
gpt-oss-120b long-context	26/18	2.33x	1.05x
Qwen3-VL-235B-FP8 image	29/26	2.08x	1.17x
Qwen3-Next-80B mixed	29/26	1.76x	1.01x
Qwen3-VL-235B-FP8 text-only	29/26	1.74x	1.05x
Kimi-Linear-48B mixed	29/26	1.69x	1.13x
Kimi-Linear-48B long-context	26/18	1.57x	0.98x
Qwen3-VL-235B-FP8 video	29/26	1.43x	1.03x
AI21-Jamba-Mini-1.7 long-context	26/18	1.14x	1.78x
Mamba-Codestral-7B mixed	26/18	1.13x	1.65x
Llama-3.1-8B long-context	26/18	1.00x	1.53x

Read top-down, the first eight rows are B200 speedups that shrink to roughly nothing on Hopper. The last three are the reverse: 0.18 deficits that exist only on Hopper.

What this means for the paper. The B200 report's two headline stories — "0.18 → 0.26 is about quantization support maturing" (carried mostly by gpt-oss-120b at 3.31x) and "0.26 → 0.29 is about quantized-MoE and linear-attention kernels" (carried by Qwen3-Next, Kimi-Linear, GLM-5.2-NVFP4 and Qwen3-VL-235B-FP8) — are Blackwell-specific. Neither survives on Hopper. They should be stated as architecture-conditional claims, not as properties of the vLLM versions.

The single claim that does transfer is the one the B200 report could not explain.

The Qwen3-VL-235B-FP8 regression reproduces to within 1%

leg	B200 image	H200 image
0.18	7,004	6,539
0.26	4,896	4,614
26/18	0.70x	0.71x

Two different architectures, two different attention backends (B200 selects FLASHINFER, H200 selects FLASH_ATTN/FA3), and the regression lands in the same place. On B200 this was measured three times per leg with under 1.5% spread. Taken together, this is a vLLM 0.26 decode-path defect for this model, and it is worth filing upstream on its own.

The text-only and video rows also match the B200 shape: text-only 0.83x here vs 0.90x there, video 0.98x vs 1.04x. The regression is concentrated in the decode-heavy workloads on both machines.

Failure attribution: H200 vs B200

model	B200 / 0.18	H200 / 0.18	verdict
whisper-large-v3	FAIL → pin	PASS	Blackwell-only
Qwen2-VL-7B	FAIL → pin	PASS	Blackwell-only
Qwen3-VL-8B	FAIL → pin	PASS	Blackwell-only
Qwen2.5-Omni-7B	FAIL → pin	PASS	Blackwell-only
Kimi-Linear-48B	FAIL → pin	PASS	Blackwell-only
Qwen3-VL-235B-FP8	FAIL → pin	PASS	Blackwell-only
GLM-5.2-NVFP4	FAIL (transformers)	SKIP (NVFP4/SM100)	n/a on Hopper
gemma-4-26B	FAIL	FAIL	genuine, architecture-independent

The cutlass failure is Blackwell-only, and this is a controlled result

The 0.18 venv on this node resolved nvidia-cutlass-dsl==4.8.0 — the same version as on B200 — and cute.core.ThrMma is genuinely absent from it, which we checked directly before starting the sweep. So the broken dependency is present and identical. All six models pass anyway.

The mechanism is backend selection. On H200 these models run FLASH_ATTN / FlashAttention 3, and the string cute, ThrMma and vllm_flash_attn/cute appear zero times in their logs — vLLM 0.18's CuTe flash-attention wrapper is never imported, so the missing symbol is never evaluated. On B200 the same models take the CuTe path and die at import.

So the upstream defect reported from the B200 round stands unchanged — vLLM 0.18.0 declares nvidia-cutlass-dsl>=4.4.0.dev1 with no upper bound while its own vllm_flash_attn/cute/utils.py:115 evaluates cute.core.ThrMma at import — but its blast radius is Blackwell only. A bug report should say so; on Hopper a fresh 0.18 install is unaffected.

gemma-4-26B: two different failures, both matching B200

Both legs fail, for the reasons the B200 report predicted, verbatim:

• 0.18 — ValidationError, does not recognize ... model type `gemma4` . transformers 4.57.6 cannot parse the config, and 0.18 contains no Gemma4 implementation either. Genuinely unsupported; not fixable.
• 0.26 — AmbiguousGlobalPerLayerAttributeError: 'head_dim' is a per-layer attribute and may vary across layers. transformers 5.17.0 tightened per-layer attribute access while vLLM 0.26 still reads config.head_dim directly. Pinning transformers==5.14.1 fixes it, giving 8,471 tok/s mixed.

Do not write this up as "0.26 does not support gemma-4". The same wording caveat as on B200.

The 0.18 long-context deficit

New on Hopper, absent on Blackwell:

model	0.18	0.26	0.29	26/18	B200 26/18
Llama-3.1-8B long-context	109	167	167	1.53x	1.00x
AI21-Jamba-Mini-1.7 long-context	65	117	117	1.78x	1.14x
Mamba-Codestral-7B mixed	4,368	7,209	7,649	1.65x	1.13x

What is controlled:

• Work done is identical. 16,384 output tokens and num_seqs=64 on every leg and both architectures. Only elapsed differs: Llama long-context is 150.1s on 0.18 against 98.3s on 0.26.
• KV cache and concurrency are identical across the three H200 legs: 890,064 / 891,392 / 879,888 tokens, 6.79x / 6.80x / 6.71x. This is lower than B200 (≈1,174,000 tokens, 8.96x) as the smaller card predicts, but it is flat across versions, so it cannot produce a 1.53x version gap.
• The attention backend is identical across the three H200 legs: all select FLASH_ATTN with FlashAttention 3. It differs from B200, where all three select FLASHINFER — but it does not differ within an architecture.
• Scheduling is identical: enable_chunked_prefill=True, max_num_batched_tokens=16384 on all three legs.

So the deficit is inside vLLM 0.18's own long-context path on the FA3 backend. 0.26 and 0.29 are within 1% of each other on these rows, so whatever it is was fixed by 0.26 and has not changed since. Isolating it further needs a kernel-level profile, which has not been done.

The practical consequence is the important part: on B200 the Llama-3.1-8B long-context row reads 300 / 300 / 300, i.e. "no version difference at all", and on H200 the same row reads 109 / 167 / 167. A reader given only the B200 table would conclude long-context performance is version-independent. It is not.

Runtime stack per leg

Identical to the B200 run, which is what makes the two comparable.

leg	vllm	torch	CUDA	transformers	flashinfer	nvidia-cutlass-dsl
0.18	0.18.0	2.10.0+cu128	12.8	4.57.6	0.6.6	4.8.0
0.26	0.26.0	2.11.0+cu130	13.0	5.17.0 → 5.14.1 (gemma-4)	0.6.14	4.6.0
0.29	0.29.0	2.13.0+cu130	13.0	5.17.0	0.6.18	4.6.2

Two notes relative to B200. The 0.29 nvidia-cutlass-dsl version, recorded as "not recorded" in the B200 manifest, is 4.6.2. And the 0.18 leg here is internally homogeneous — no cutlass pin was needed — so only the 0.26 leg carries a per-model dependency difference, and only for gemma-4.

Known limitations

1. No dispersion. Every throughput number is a single timed pass. Treat anything in the 0.97–1.05x band as "no difference detected", not "no difference". The B200 round repeated Qwen3-VL-235B-FP8 three times per leg and found under 1.5% spread, which is encouraging for the rest but was not re-done here.
2. H200 only, and not a hardware benchmark. These numbers are for comparing vLLM versions on Hopper. The B200 and H200 absolute figures should not be put in one table as a hardware comparison: the two machines select different attention backends, and image/video throughput additionally depends on the CPU allocation the harness hands each job.
3. GLM-5.2-NVFP4 has no data on Hopper at all, so the B200 finding for it (2.47x on 0.26 → 0.29) has no counterpart here and cannot be checked.
4. One dependency pin on the 0.26 leg, for gemma-4 only. The other 13 models on that leg were measured under transformers 5.17.0. Disclose rather than reporting a single version for the leg.
5. The long-context deficit is characterised, not explained. See above.

Where the numbers come from

The tables are generated from results/ by make_tables.py, not transcribed:

python3 make_tables.py   # run from the directory containing results/

compare/*.txt is produced from the same canonical run set, and README.md is generated from this file by make_card.py. CANONICAL.md states which run is authoritative for every (version × model) cell.