This is the first environment-clean dataset, and it is citable. Every performance number from the 2026-09-20 round is void: the 0.26 leg was contaminated by the vllm-omni plugin, and all jobs ran concurrently. All three legs in this report use isolated venvs, run strictly serially with the whole node to themselves, and were measured without --resume, so every number was actually measured. But there is still almost no repeated measurement — throughput is a single timed pass except for one model; 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 H200 (Hopper) is at Alexsssu/fastkernels-h200-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 conclusions below do not reproduce there; each headline bullet is annotated with its H200 counterpart.

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/`	49 `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`	58 `run.log`, one per task
`patches/`	the uncommitted local changes, the scenario tables, the driver scripts
`moe_backend_235b/`, `diag_silent4_summary.txt`	the two diagnostic sub-studies
`SOURCE_NOTES.md`	internal working documents deliberately not shipped, and why
`raw/`	raw harness output (194 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	failing
0.18.0	13/15	`GLM-5.2-NVFP4`, `gemma-4-26B`
0.26.0	15/15	—
0.29.0	15/15	—

13 models support a three-way comparison (the clean re-run alone gave only 7). The 6 that came back all did so from a single dependency pin, see "Failure attribution". The 2 remaining 0.18 failures look identical from the outside, but only gemma-4-26B is genuinely a case of "this older version does not support this architecture".

Headline findings

Read this first. A companion run of this exact ablation on 8×H200 (2026-09-22, same scenario table, same tp, same three runtime stacks) shows that most of the findings below are Blackwell-specific and do not reproduce on Hopper. Each bullet is annotated with its H200 counterpart. Treat them as architecture-conditional claims, not as properties of the vLLM versions.

0.18 → 0.26 is a story about whether a model runs at all, and about quantization support maturity — not about a general speedup. Plain dense bf16 Llama-3.1-8B lands within 1% across all three legs (21,943 / 22,263 / 22,157 tok/s). The single large jump is gpt-oss-120b at 3.31x (7,304 → 24,184), which corresponds to mxfp4 support maturing. On H200 this is 1.03x. vLLM 0.18 already reaches 11,290 tok/s there, so the 3.31x measures a Blackwell-specific deficiency in 0.18 rather than an improvement in 0.26.
0.26 → 0.29 is a story about quantized-MoE and linear-attention kernels. The winners are GLM-5.2-NVFP4 at 2.47x, Qwen3-VL-235B-FP8 image at 2.08x, Qwen3-Next-80B at 1.76x and Kimi-Linear-48B at 1.69x. Dense, ASR and the multimodal encoder paths are essentially flat (0.99–1.05x). On H200 this story disappears: 235B image 1.17x, Qwen3-Next-80B 1.01x, Kimi-Linear-48B 1.13x, and GLM-5.2-NVFP4 cannot run at all because NVFP4 is SM100-only. 0.26 → 0.29 is close to flat across the board on Hopper.
Version progress is not monotonic. Qwen3-VL-235B-FP8 is slower on 0.26 than on 0.18 (image 7,004 → 4,896, i.e. 0.70x) and only recovers to 10,175 on 0.29. This one has since been measured three times per leg and it holds; the cause is still open, see "Is the 0.26 regression real?". Qwen3-VL-8B is flat across all three (15,582 / 15,496 / 15,336), with 0.29 marginally lowest. Do not tell this as "newer is faster". This is the one finding that does transfer: H200 measures 6,539 → 4,614, i.e. 0.71x against 0.70x here, on a different architecture with a different attention backend. That agreement makes a real vLLM 0.26 decode-path defect the most likely explanation.

The three-way agreement on Llama-3.1-8B is the most important cross-check in this round: it positively confirms that the 3x dip in the old data was an artifact of the vllm-omni plugin plus concurrent neighbours, not a vLLM version difference. The old claim of a "3.03x speedup from 0.26 to 0.29" does not hold.

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	21,943	22,263	22,157	1.01x	1.00x
`Llama-3.1-8B-Instruct`	1	long-context	300	300	300	1.00x	1.00x
`Mixtral-8x7B-Instruct-v0.1`	2	mixed	14,177	15,351	18,740	1.08x	1.22x
`Mixtral-8x7B-Instruct-v0.1`	2	long-context	549	581	579	1.06x	1.00x
`gpt-oss-120b`	2	mixed	7,304	24,184	25,478	3.31x	1.05x
`gpt-oss-120b`	2	long-context	277	644	651	2.33x	1.01x
`mamba-2.8b-hf`	1	mixed	8,572	8,889	10,298	1.04x	1.16x
`mamba-2.8b-hf`	1	long-context	331	362	362	1.10x	1.00x
`Mamba-Codestral-7B-v0.1`	1	mixed	10,333	11,640	11,775	1.13x	1.01x
`Mamba-Codestral-7B-v0.1`	1	long-context	481	490	497	1.02x	1.01x
`Qwen3-Next-80B-A3B-Instruct`	2	mixed	10,201	10,329	18,227	1.01x	1.76x
`Qwen3-Next-80B-A3B-Instruct`	2	long-context	589	749	807	1.27x	1.08x
`AI21-Jamba-Mini-1.7`	1	mixed	4,404	5,498	5,552	1.25x	1.01x
`AI21-Jamba-Mini-1.7`	1	long-context	184	209	213	1.14x	1.02x
`Kimi-Linear-48B-A3B-Instruct`	2	mixed	13,224	13,550	22,910	1.02x	1.69x
`Kimi-Linear-48B-A3B-Instruct`	2	long-context	563	885	903	1.57x	1.02x
`whisper-large-v3`	1	librispeech	7,888	8,008	8,135	1.02x	1.02x
`Qwen2-VL-7B-Instruct`	1	text-only	16,079	17,482	17,620	1.09x	1.01x
`Qwen2-VL-7B-Instruct`	1	image	14,424	16,247	16,837	1.13x	1.04x
`Qwen2-VL-7B-Instruct`	1	video	2,487	2,600	3,082	1.05x	1.19x
`Qwen3-VL-8B-Instruct`	1	text-only	13,236	13,772	13,913	1.04x	1.01x
`Qwen3-VL-8B-Instruct`	1	image	15,582	15,496	15,336	0.99x	0.99x
`Qwen3-VL-8B-Instruct`	1	video	2,656	2,612	2,702	0.98x	1.03x
`Qwen3-VL-235B-A22B-Instruct-FP8`	4	text-only	4,444	3,985	6,934	0.90x	1.74x
`Qwen3-VL-235B-A22B-Instruct-FP8`	4	image	7,004	4,896	10,175	0.70x	2.08x
`Qwen3-VL-235B-A22B-Instruct-FP8`	4	video	1,390	1,451	2,075	1.04x	1.43x
`Qwen2.5-Omni-7B`	1	text	16,254	17,532	17,576	1.08x	1.00x
`Qwen2.5-Omni-7B`	1	image	13,120	13,627	13,957	1.04x	1.02x
`Qwen2.5-Omni-7B`	1	video	2,120	2,105	2,166	0.99x	1.03x
`Qwen2.5-Omni-7B`	1	audio	4,851	6,746	6,836	1.39x	1.01x
`GLM-5.2-NVFP4`	8	mixed	—	4,306	10,617	—	2.47x
`GLM-5.2-NVFP4`	8	long-context	—	248	267	—	1.08x
`gemma-4-26B-A4B-it`	1	mixed	—	10,528	12,171	—	1.16x
`gemma-4-26B-A4B-it`	1	long-context	—	100	117	—	1.17x

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	3.82	3.73	3.83	1.02x	0.97x
`Llama-3.1-8B-Instruct`	1	fixed-batch-32	0.18	0.17	0.18	1.04x	0.99x
`Mixtral-8x7B-Instruct-v0.1`	2	single-request	4.81	3.76	3.52	1.28x	1.07x
`Mixtral-8x7B-Instruct-v0.1`	2	fixed-batch-32	0.38	0.33	0.33	1.15x	1.00x
`gpt-oss-120b`	2	single-request	4.04	2.45	2.50	1.65x	0.98x
`gpt-oss-120b`	2	fixed-batch-32	0.49	0.19	0.19	2.51x	1.01x
`mamba-2.8b-hf`	1	single-request	3.61	3.16	3.20	1.14x	0.99x
`mamba-2.8b-hf`	1	fixed-batch-32	0.37	0.35	0.35	1.05x	1.00x
`Mamba-Codestral-7B-v0.1`	1	single-request	5.01	4.02	3.74	1.25x	1.07x
`Mamba-Codestral-7B-v0.1`	1	fixed-batch-32	0.27	0.23	0.23	1.15x	1.03x
`Qwen3-Next-80B-A3B-Instruct`	2	single-request	5.88	4.31	3.56	1.36x	1.21x
`Qwen3-Next-80B-A3B-Instruct`	2	fixed-batch-32	0.36	0.28	0.27	1.31x	1.03x
`AI21-Jamba-Mini-1.7`	1	single-request	7.53	6.93	6.93	1.09x	1.00x
`AI21-Jamba-Mini-1.7`	1	fixed-batch-32	0.86	0.82	0.81	1.05x	1.01x
`Kimi-Linear-48B-A3B-Instruct`	2	single-request	4.62	3.75	2.48	1.23x	1.51x
`Kimi-Linear-48B-A3B-Instruct`	2	fixed-batch-32	0.29	0.24	0.22	1.19x	1.11x
`whisper-large-v3`	1	single-utterance	2.63	2.35	2.49	1.12x	0.95x
`whisper-large-v3`	1	fixed-batch-32	0.18	0.17	0.17	1.02x	0.98x
`Qwen2-VL-7B-Instruct`	1	single-image	3.83	3.82	3.77	1.00x	1.01x
`Qwen2-VL-7B-Instruct`	1	single-video	5.20	5.12	4.76	1.02x	1.08x
`Qwen3-VL-8B-Instruct`	1	single-image	4.66	4.27	4.22	1.09x	1.01x
`Qwen3-VL-8B-Instruct`	1	single-video	5.79	5.44	5.30	1.06x	1.03x
`Qwen3-VL-235B-A22B-Instruct-FP8`	4	single-image	10.21	8.23	7.56	1.24x	1.09x
`Qwen3-VL-235B-A22B-Instruct-FP8`	4	single-video	12.30	10.24	8.92	1.20x	1.15x
`Qwen2.5-Omni-7B`	1	single-text	3.67	3.65	3.61	1.01x	1.01x
`Qwen2.5-Omni-7B`	1	single-image	3.92	3.90	3.88	1.01x	1.00x
`Qwen2.5-Omni-7B`	1	single-video	5.46	5.47	5.41	1.00x	1.01x
`Qwen2.5-Omni-7B`	1	single-audio	3.90	3.97	3.96	0.98x	1.00x
`GLM-5.2-NVFP4`	8	single-request	—	8.98	6.97	—	1.29x
`GLM-5.2-NVFP4`	8	fixed-batch-32	—	0.50	0.46	—	1.09x
`gemma-4-26B-A4B-it`	1	single-request	—	4.34	4.10	—	1.06x
`gemma-4-26B-A4B-it`	1	fixed-batch-32	—	0.40	0.38	—	1.06x

Failure attribution

Two root causes, not three

The targeted backfill on 2026-09-21 overturned the earlier three-way split. Our runbook had grouped the 0.18 leg's 8 failures as A (architecture not recognized), B (cutlass API drift) and C (multimodal encoder dying silently). B and C turned out to be the same bug, and a fixable one.

root cause	models	error	fixable?
A · config not parseable by this leg's transformers	2	pydantic `ValidationError` while constructing `ModelConfig`: model type `glm_moe_dsa` / `gemma4` not recognized	No, not within 0.18's declared constraints
B · cutlass `ThrMma` relocation	6	`AttributeError: module 'cutlass.cute.core' has no attribute 'ThrMma'`	Yes — pin `nvidia-cutlass-dsl==4.5.3`

Group B is Kimi-Linear-48B and Qwen3-VL-235B-FP8 (originally class B) plus whisper-large-v3, Qwen2-VL-7B, Qwen3-VL-8B and Qwen2.5-Omni-7B (originally class C). All six now pass and have real numbers.

So the number of models that can be described as "this older version does not support this architecture" is 1 (gemma-4-26B), not the 8 we originally assumed. The other 7 are all dependency-version problems: 6 are fully recovered by a single cutlass pin, and GLM-5.2-NVFP4 is a different flavour of version-window problem, covered next.

Group A: only one of the two is genuinely unsupported

Both models fail identically from the outside, so it is tempting to write both off as "vLLM 0.18 does not support this architecture". Only one of them is.

gemma-4-26B genuinely is unsupported. It declares Gemma4ForConditionalGeneration / model_type: gemma4, and vLLM 0.18 contains no Gemma4 code at all — grep -rn Gemma4 vllm/model_executor/models/ returns nothing, and the registry stops at Gemma3nForConditionalGeneration. Even with a transformers that could parse the config, there would be no implementation to run.
GLM-5.2-NVFP4 is a version-window problem, not missing support. The model declares GlmMoeDsaForCausalLM / glm_moe_dsa, and vLLM 0.18 implements exactly that: registry.py:119 maps GlmMoeDsaForCausalLM, implemented at deepseek_v2.py:1638. What blocks it is transformers: the error text says Transformers does not recognize the architecture, not vLLM. glm_moe_dsa first appears in transformers 5.x (absent in 4.57.6, present in 5.17.0), while vLLM 0.18 declares transformers<5,>=4.56.0. Since 4.57.6 is the final 4.x release — transformers went straight to 5.0.0 — no version inside that constraint can parse this checkpoint.

So vLLM 0.18 ships an architecture whose config it can never parse under its own dependency constraints. Structurally this is the same kind of version-window defect as the cutlass one below; the difference is that the cutlass window (>=4.4.0.dev1, <4.6.0) falls inside what vLLM declares, so a pin fixes it, whereas the transformers window for glm_moe_dsa (>=5) falls entirely outside.

Wording to use externally: say "0.18 does not support this architecture" only for gemma-4-26B. For GLM-5.2-NVFP4, say that 0.18 implements the architecture but cannot be paired with a transformers able to parse the checkpoint.

One assumption here is inferred rather than measured: that vLLM 0.18 cannot run against transformers 5.x at all. That rests on the declared <5 bound, not on an experiment. Force-installing transformers 5.x into the 0.18 venv and retrying GLM-5.2-NVFP4 would settle it in minutes. Prior is that it breaks — transformers 5.17 already broke vLLM 0.26's config.head_dim access (the gemma-4 failure below), and 0.18 is older.

Why group C looked like a silent crash

All four models stopped on the same line during the sweep,

Encoder cache will be initialized with a budget of 16384 tokens

then exited rc=1 with no traceback. The crash was not actually silent: the error is raised inside the EngineCore subprocess, and the sweep deliberately runs without DEBUG logging (compare_vllm_versions_datafast.sh:61). DEBUG output corrupts the stdout of the cpuinfo subprocess that vLLM's usage-reporting thread json-parses; the resulting thread exception prints a traceback into the worker log, and ray_runner._is_fatal_worker_line() treats that as a crash and kills a healthy worker — which is how one attempt at this sweep lost the entire 0.18 leg.

Re-running outside the harness with DEBUG on, diag_silent4_018.sh surfaced the traceback immediately:

AttributeError: module 'cutlass.cute.core' has no attribute 'ThrMma'
RuntimeError: Engine core initialization failed

Word for word the group B error.

The cutlass chain, end to end

vLLM 0.18 bundles its own CuTe flash-attention wrapper, vllm/vllm_flash_attn/cute/utils.py, whose line 115 carries the module-level annotation thr_mma: cute.core.ThrMma — evaluated at import time.
nvidia-cutlass-dsl moved ThrMma out of cute.core into cute.atom in 4.6.0 (still exported as cutlass.cute.ThrMma). Measured directly: 4.3.4 / 4.4.0 / 4.5.3 have it under cute.core; 4.6.0 / 4.7.1 / 4.8.0 do not.
vLLM 0.18 requires nvidia-cutlass-dsl>=4.4.0.dev1 and flashinfer 0.6.6 requires >=4.3.4 — neither declares an upper bound, so uv resolves 4.8.0.
A fresh 0.18 install therefore cannot import its own flash-attention wrapper. The working window is >=4.4.0.dev1, <4.6.0, and 4.5.3 is the newest in it.

This is directly reportable upstream: vLLM 0.18.0's dependency metadata admits a range of nvidia-cutlass-dsl in which vLLM 0.18.0 cannot work.

The diag_silent4_018.sh control is single-variable — four falsification arms all fail, and only changing the cutlass version makes all four models pass (full output in diag_silent4_summary.txt):

arm	whisper	Qwen2-VL	Qwen3-VL-8B	Omni
baseline (cutlass 4.8.0)	rc=1	rc=1	rc=1	rc=1
`--enforce-eager`	rc=1	rc=1	rc=1	rc=1
`VLLM_ATTENTION_BACKEND=TRITON_ATTN`	rc=1	rc=1	rc=1	rc=1
`--gpu-memory-utilization 0.5`	rc=1	rc=1	rc=1	rc=1
cutlass-dsl 4.5.3	rc=0	rc=0	rc=0	rc=0
0.26 control	rc=0	rc=0	rc=0	rc=0

gemma-4 on the 0.26 leg: transformers' fault, not vLLM's

AmbiguousGlobalPerLayerAttributeError: 'head_dim' is a per-layer attribute and
may vary across layers. Access it via the individual layer configs instead

transformers 5.17.0 tightened per-layer attribute access, while vLLM 0.26's model_arch_config_convertor.py:597 still reads config.head_dim directly. Pinning transformers==5.14.1 — the version pyproject.toml declares and the one validated against 0.26 — makes it run, at 10,528 tok/s mixed. Do not write this up as "0.26 does not support gemma-4".

Is the 0.26 regression on Qwen3-VL-235B-FP8 real?

Yes, and the cause is still unknown. This is the only non-monotonic result in the table, so it was the first thing re-measured.

Three repeats per leg: it holds

scenarios_repeat235b.yaml with REPS=3, both legs, repetition as the outer loop so slow drift cannot line up with the version axis. Archived under results/vllm-*-vllm-only-rep235-r[123].

workload	0.18 (3 reps)	median	0.26 (3 reps)	median	26/18
text-only	4,503 / 4,512 / 4,482	4,503	3,947 / 3,956 / 3,953	3,953	0.88x
image	7,030 / 6,985 / 6,948	6,985	4,862 / 4,919 / 5,049	4,919	0.70x
video	1,406 / 1,460 / 1,410	1,410	1,422 / 1,450 / 1,478	1,450	1.03x

Repeat spread is under 1.5% on both legs while the image gap is 42%, so this is signal, not noise. Both medians also land within ~1% of the single-pass values in the main table (0.18 image 6,985 vs 7,004; 0.26 image 4,919 vs 4,896), which is why the table was left as it is.

The MoE kernel is not the explanation

Both legs log the same nominal MoE backend (FLASHINFER_TRTLLM Fp8), but 0.26 routes it differently:

	prepare/finalize	SM split
0.18	`MoEPrepareAndFinalizeNoDPEPModular`	none
0.26	`MoEPrepareAndFinalizeNoDPEPMonolithic`	`140 SMs used for MoE, 8 reserved`

That looked like the answer, so --moe-backend was added to bench_vllm.py (it is part of the phase fingerprint, so --resume cannot reuse a run measured under a different backend) and the alternatives were forced. Results in moe_backend_235b/summary.txt:

arm	image tok/s	prepare/finalize	SM split
0.18 auto	7,457	Modular	none
0.26 auto (resolves to trtllm)	5,536	Monolithic	140 SMs
0.26 `flashinfer_trtllm`	5,614	Monolithic	140 SMs
0.26 `triton`	5,782	Modular	none
0.26 `flashinfer_cutlass`	refused	—	—
0.26 `cutlass`	refused	—	—

flashinfer_cutlass and cutlass refuse this model legitimately — does not support quantization scheme QuantKey(f8e4m3fn...) — not a bug. The same arm was also run on 0.18 (018_flashinfer_cutlass, rc=1) and refuses for the same reason; it is in moe_backend_235b/summary.txt but carries no number, so it is not in the table above.

The decisive negative result is triton. It restores exactly the structure 0.18 has (Modular prepare/finalize, no SM reservation) and still only reaches 5,782, nowhere near 7,457. So the Monolithic prepare/finalize and the SM reservation are not sufficient to explain the gap. This is not airtight — triton is a slower kernel in its own right, so the arm changes two things at once — but no MoE backend available on 0.26 recovers 0.18's throughput.

Ruled out so far

Work done: identical. Same num_seqs (1000), same vllm_output_tokens (417,061 text-only / 512,000 image and video), same output_len, tp, seed. Only elapsed differs.
Attention backend: both legs pick FLASH_ATTN for MMEncoderAttention.
Scheduling: both have chunked prefill on with max_num_batched_tokens=16384.
KV cache and concurrency: 2,173,456 tokens / 128.64x on 0.26 and 2,192,112 / 129.74x on 0.18 — a 0.9% difference, not 42%.
MoE backend selection: see above.
FlashInfer autotuning landing inside the timed region on one leg only: it runs fresh on every launch and is paid equally by both. Each arm tunes the full 22 flashinfer::trtllm_fp8_block_scale_moe configurations (~~7–8s), and nothing persists between runs — `~~/.cache/flashinfer/{0.6.6,0.6.14}` was last written hours before these runs and was not touched during them. So it is a constant overhead on every measurement, not a version-dependent one.

The remaining lead

The regression tracks how decode-bound the workload is: image (decode-heavy) 0.70x, text-only (pure decode) 0.88x, video (prefill- and encoder-heavy, ~1,400 tok/s) 1.03x. That points at 0.26's decode path for this model rather than at the vision encoder. Settling it needs a kernel-level profile of the image workload on both legs; that has not been done.

Runtime stack per leg

A "vLLM version comparison" is really a comparison of three whole runtime stacks, so this table has to travel with the numbers.

leg	vllm	torch	CUDA	transformers	nvidia-cutlass-dsl
0.18	0.18.0	2.10.0+cu128	12.8	4.57.6	4.8.0 → 4.5.3 (6 models)
0.26	0.26.0	2.11.0+cu130	13.0	5.17.0 → 5.14.1 (gemma-4)	4.6.0
0.29	0.29.0	2.13.0+cu130	13.0	5.17.0	not recorded

Why the pins did not contaminate the existing numbers

The 0.18 leg spans two cutlass configurations: 7 models measured under 4.8.0 and 6 under 4.5.3. This does not affect the comparability of the first 7, and the argument is deterministic: they already passed under 4.8.0, which proves they never reach the vllm_flash_attn/cute code path — reaching it fails at import — so the cutlass version cannot have changed their measurements. Likewise transformers==5.14.1 was used only for gemma-4; the other 14 numbers on the 0.26 leg were still measured under 5.17.0.

The stricter alternative is to re-measure each leg whole under one pinned set, at a cost of several GPU hours. The argument above holds, so that was not done. The inconsistency still has to be disclosed — do not report a single cutlass or transformers version per leg.

Known limitations

Almost no dispersion, and this is the biggest gap. Throughput is one timed pass per scenario for 14 of the 15 models; REPS=3 has only been run for Qwen3-VL-235B-FP8. We cannot currently answer "is the difference between 22,263 and 21,943 real?", so every conclusion in the 0.99–1.05x band should be read as "no difference detected", not "no difference". This most affects the "flat across all three" claims for Qwen3-VL-8B and whisper. The one model that was repeated came back with under 1.5% spread, which is encouraging for the rest but does not substitute for measuring them. Note that scenarios_repeat7.yaml is now stale: 13 models are three-way comparable, not 7, so repeated measurement needs either a 13-model table or an explicit statement that dispersion covers only some of them.
Image and video throughput depends on the CPU allocation the harness hands the job, so these absolute numbers are not portable. With mm_processor_cache_gb=0 the multimodal preprocessing is re-run inside the timed region, which makes those workloads partly CPU-bound. Under the Ray harness a tp=4 job gets OMP_NUM_THREADS=21 (ray_runner.py:499, per_rank = num_cpus // tp, with num_cpus allocated tp-proportionally out of 192 cores). Running the same benchmark directly, without Ray, leaves the variable unset and each rank takes all 192 cores — which moved Qwen3-VL-235B-FP8 image by 6–12% (0.26: 105.3s → 92.5s; 0.18: 72.8s → 68.7s) while leaving text-only unchanged, exactly as a CPU-bound preprocessing step would. Version-to-version comparisons in this report are unaffected, since every leg ran through the same harness at the same tp, but do not compare these image/video figures against numbers produced by a different runner. Counter-intuitively the direct runs are also much noisier (17% between two identically configured arms, against ~1% in the harness), most likely because 4 ranks × 192 threads oversubscribes 192 cores; the Ray-imposed 21 threads is the better-behaved configuration.
Whisper audio is clamped to 30s. LibriSpeech test.clean does contain samples longer than that (longest ≈30.61s). The clamp works around the vLLM 0.29 whisper bug, and Whisper's encoder is defined on a fixed 30s window anyway, so it does not change the semantics of the workload — but it must be stated explicitly.
B200 only, and most conclusions do not generalise. Do not put these absolute numbers in the same table as Hopper results; reasoning in MANIFEST.md. Beyond that, the H200 companion run shows that the two headline speedup stories above are Blackwell-specific — see the note at the top of "Headline findings". The 0.18 failure set is also Blackwell-specific: on Hopper 0.18 reaches 13/14 with no dependency pins at all.
Two dependency pins leave two legs internally non-homogeneous. See "Why the pins did not contaminate the existing numbers". The argument holds but must be disclosed.
gemma-4 has two-way data only (pinned 0.26 plus 0.29); 0.18 cannot load it, so it carries no three-way comparison. Its long-context figure of 117 tok/s is the lowest in the table, but it is the largest model running at tp1 (26B), and being 2.5x slower than Llama-3.1-8B at tp1 (300) for the same amount of work is reasonable, not an outlier.

Three defects found along the way (independent of the performance numbers)

vLLM 0.18.0's dependency metadata admits an nvidia-cutlass-dsl range in which it cannot work. 0.18 declares >=4.4.0.dev1 with no upper bound, but its own vllm_flash_attn/cute/utils.py:115 evaluates cute.core.ThrMma at import, and that symbol left cute.core in cutlass-dsl 4.6.0. The real window is >=4.4.0.dev1, <4.6.0. Consequence: a fresh 0.18 install today resolves 4.8.0 and six models (Kimi-Linear-48B, Qwen3-VL-235B-FP8, whisper-large-v3, Qwen2-VL-7B, Qwen3-VL-8B, Qwen2.5-Omni-7B) all die during engine initialization. Pinning 4.5.3 makes all six pass — the single-variable table in diag_silent4_018.sh is ready-made evidence. The blast radius is Blackwell only, and a bug report should say so. On H200 the same 0.18 venv resolves the same broken 4.8.0, and all six models pass unpinned: Hopper selects FLASH_ATTN/FA3, so the CuTe wrapper carrying the bad symbol is never imported. The metadata defect is real either way, but it is only reachable on SM100.
A whisper bug in vLLM 0.29. For one ≈30.61s librispeech clip, 0.29 does not truncate to Whisper's standard 3000 frames and produces 3,061 mel frames; its own multimodal batching code then refuses to stack 3,061 with 3,000: input_features contains inconsistent shapes. Our harness hands vLLM raw audio and lets vLLM do its own feature extraction; running the same WhisperProcessor over the whole split independently yields (128, 3000) for all 2,620 samples. The input is entirely valid, and 0.26 passes in the same isolated venv. Ready to file upstream.
An IpcSocket defect in our own bench_vllm.py (worth upstreaming as a PR). Before the fix, every tp>1 job on 0.29 died within 18–37 seconds on AttributeError: module 'flashinfer.comm.mnnvl' has no attribute 'IpcSocket'. After it, Mixtral (tp2), gpt-oss-120b (tp2), Qwen3-Next-80B (tp2), Kimi-Linear-48B (tp2), Qwen3-VL-235B-FP8 (tp4) and GLM-5.2-NVFP4 (tp8) all pass. This round's data is the before/after evidence.

Where the numbers come from

The tables are generated from results/ by make_tables.py, not transcribed, so they cannot drift away from the data:

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

make_tables.py reads exactly one run directory per (leg, model) — the canonical measurement. Which directory that is, and why, is listed model by model in CANONICAL.md. The repeat and MoE-backend runs that also live under results/ are deliberately not folded into these tables; they are supporting evidence for the two diagnostic sections above.

Downloads last month: 25

Total file size:

205 MB