[zcnrex]

For Qwen3-VL-4B text traffic on one B300, document a BF16 mixed-chunk recipe and its measured latency tradeoff on a pinned source build. The matched seven-point sweeps use the same source build, BF16 weights/KV cache/queries, TRTLLM attention, page size 32, 16384-token chunks, and the default NHD layout; only mixed chunking changes.

At concurrency 128, median TTFT falls from 4120 to 578 ms, output throughput changes from 3764 to 3802 tokens/s, and median TPOT rises from 29.97 to 32.95 ms. The page includes the full comparison, benchmark protocol, and source-build prerequisite in #42913. Input length 8192 is nominal before chat templating and retokenization; actual model input lengths vary. The guidance asks users to remeasure mean and tail latency as well as medians after changing runtime configuration. Text GSM8K scores 1216/1314 versus 1215/1314 for the reference; a one-image smoke test produced the same 32 generated tokens, but broad image/video accuracy was not evaluated.

Raw measurements and reproduction scripts: https://github.com/zcnrex/sglang/tree/f907431182d8097b1317d3029f006fb55b8e23df/benchmark/qwen3_vl

A separate accordion records the HND configuration at exact code commit c76cfaba1a1154a4a24b8694184cb8d3318fc358, including the pinned model, explicit 20 decode graph buckets through 128, 1,600,000-token KV capacity, and seeds. It reports output-only throughput and median TTFT/TPOT across seven concurrency points plus a second C128 GPU, with C128 mean/p95/p99 TTFT. This snapshot is explicitly separate from the historical NHD off/on comparison: concurrent GPUs and normal public autotuning varied, so it does not establish a causal layout or mixed-chunk gain. Separate C1 GSM8K validation scores1217/1314 in both arms; C2 scores1213/1314 with the small-batch extension versus1216/1314 without it despite matching startup tactics. The page retains this deficit and the scope of a four-prompt512-step fixed-input bitwise diagnostic; it makes no broad accuracy-equivalence claim.

Pinned HND evidence: https://github.com/zcnrex/sglang/tree/a2f3a4e381d359e3fb4f1e9984ceb9f94b6a3f2e/benchmark/qwen3_vl/results/final_qkv_full_sweep

Validation: mint validate, mint broken-links, targeted pre-commit hooks, and git diff --check passed.

---

CI States

Latest PR Test (Base): ✅ Run #37744121571
Latest PR Test (Extra): ❌ Run #37744121152
Latest PR Test (AMD ROCm 10): ➖ No AMD PR run found for this commit.