benchmark: add Qwen3-VL vision-encoder (ViT) SDPA config + GB300 results

[vedaanta]

What

Adds a sdpa_benchmark_training config for the Qwen3-VL vision encoder (ViT) self-attention as exercised by image/video inference, plus GB300 results (CSV + chart).

• Model preset: 16 MHA heads, head_dim 72 zero-padded to 80 (the 16-byte-alignment contract fp8 kernels require and production integrations run; reported TFLOPS count d=80).
• Mask/pass: bidirectional (no_mask), forward-only, batch 1 — ViT inference.
• Sequence lengths: six real per-image patch-grid token counts from a production inference trace (8836 = 94², 15376 = 124², 24336 = 156², 35344 = 188², 47376, 62500 = 250²). They span the trace's per-forward FLOPs distribution from the 10th to the 99th percentile; single-image (batch-1) forwards dominate that workload. 15376 is the most frequent forward, 35344 the FLOPs-median.

Results (GB300, cuDNN backend 9.25.0, FE 1.27.0)

S	bf16	fp8
8836	1224 TF	1537 TF
15376	1247 TF	1639 TF
24336	1249 TF	1694 TF
35344	1245 TF	1738 TF
47376	1246 TF	1718 TF
62500	1264 TF	1675 TF

fp8 is a consistent ~1.35x over bf16 across the whole range.

flash_attention_4 rows are recorded as failed in the CSV: the released fa4 4.0.0b3 wheel's SM100 forward asserts Only SM 10.x and 11.x are supported on this part (sm103 missing from its arch table). Can be revisited when a newer fa4 wheel lands.

Testing

• runner --config qwen3vl_vit --dry-run expands to the expected 18 cases.
• Full run on GB300 (NGC pytorch:26.03 + latest FE/backend wheels): 12/12 cuDNN cases pass; CSV + qwen3vl_vit_no_mask.png generated by the suite's chart path.

🤖 Generated with Claude Code

Summary by CodeRabbit

• New Features

  • Added a benchmark configuration for the Qwen3-VL vision encoder, including production-derived sequence lengths and supported attention backends.
  • Added benchmarking support for bfloat16 and FP8 data types with forward-only profiling.

• Documentation

  • Added instructions for running the Qwen3-VL vision-encoder benchmark.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: ee29db60-bcc1-458b-9440-b47f81cc6b26

📥 Commits

Reviewing files that changed from the base of the PR and between 8a3b867 and d6de843.

⛔ Files ignored due to path filters (2)

• benchmark/sdpa_benchmark_training/results/qwen3vl_vit/gb300/qwen3vl_vit_20260814_193600.csv is excluded by !**/*.csv
• benchmark/sdpa_benchmark_training/results/qwen3vl_vit/gb300/qwen3vl_vit_no_mask.png is excluded by !**/*.png

📒 Files selected for processing (2)

• benchmark/sdpa_benchmark_training/README.md
• benchmark/sdpa_benchmark_training/configs/qwen3vl_vit.py

🚧 Files skipped from review as they are similar to previous changes (2)

• benchmark/sdpa_benchmark_training/configs/qwen3vl_vit.py
• benchmark/sdpa_benchmark_training/README.md

---

📝 Walkthrough

Walkthrough

Added a Qwen3-VL vision-encoder SDPA benchmark configuration with production-derived settings, supported execution backends and data types, and README usage documentation.

Changes

Qwen3-VL ViT benchmark

Layer / File(s)	Summary
Qwen3-VL ViT preset and usage benchmark/sdpa_benchmark_training/configs/qwen3vl_vit.py, benchmark/sdpa_benchmark_training/README.md	Added the QWEN3VL_VIT and CONFIG definitions. The preset uses 16 query/KV heads, an 80-dimensional head, six sequence lengths, cuDNN and FlashAttention 4, bfloat16 and FP8, unmasked forward-only profiling, batch size 1, and configured warmup and iteration counts. The README lists the configuration and its runner command.

Estimated code review effort: 2 (Simple) | ~10 minutes

Merge Risk: ⚪ Minimal · up to d6de8

This change adds a localized benchmark configuration and GB300 results without evidence of a merge-blocking correctness, runtime, deployment, or availability risk; it is merge-ready after normal checks and review.

Suggested labels: mod-frost

Suggested reviewers: brandonfzhang

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Description check	⚠️ Warning	The description explains the benchmark configuration, results, limitations, and testing, but omits several template sections and checklist items.	Add the required checklist, affected area, explicit Why, Related issues, and API and compatibility impact sections.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly identifies the added Qwen3-VL ViT SDPA configuration and associated GB300 benchmark results.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[vedaanta]

Refreshed after #597 merged (thanks @brandon): backends now cudnn / cudnn_oss / flash_attention_4, GB300 results regenerated in one session (FE built from develop tip, backend 9.25).

S	cudnn BF16	FAv4 BF16	cudnn FP8
8836	1215	1457	1547
15376	1241	1492	1629
24336	1245	1476	1707
35344	1241	1491	1721
47376	1253	1504	1709
62500	1269	1423	1648

• FA4 works on GB300 via pip install --pre "flash-attn-4[cu13]" — the released fa4 4.0.0b3 wheel asserts sm103 unsupported, the pre-release wheel does not. FAv4 BF16 runs ~18% ahead of the cudnn BF16 rows on these shapes; cudnn FP8 tops all series.
• cudnn_oss rows are recorded unsupported (exit-42 path): no FROST OSS engine serves a bare-dense fwd graph whose S_kv isn't a multiple of 128 (unmasked-tail rule), and ViT patch-grid lengths (94², 124², …) never are. Engine-side dense tail handling would light these rows up — noting it as the concrete follow-up this workload motivates.