[Levius-Fubuki]

Required human contributor answers

The contributor explicitly reviewed and confirmed these answers and PR necessity in the implementation task on 2026-10-25.

1. What outcome are you trying to achieve?

   Enable language CUDA Graphs when Cua-S1 vision features stay on the GPU, preserving screenshot preparation, prompts, probabilities and decision semantics. The goal is lower repeated-decision latency where measurements justify it; capture cost and fallback remain explicit.

2. What is the highest-leverage variable for that outcome?

   CPU operator submission and kernel-launch overhead in language forward. The primary outcome is warmed complete-request p50/p95, with first-capture cost and GPU memory reported separately. Correct current-input execution is a prerequisite.

3. How will this work move that variable rather than create busywork?

   Combine the existing resident-feature path and existing multimodal Graph path by staging current input into owned scratch before replay. Use matched eager/Graph evidence; keep the opt-in and do not claim every shape improves.

4. Who is the customer or user, and what do they actually need?

   GUI/browser-agent operators serving Cua-S1 and native-worker maintainers. They need predictable screenshot decisions with unchanged API behavior, current data after screenshot/shape changes, safe cache retirement and recoverable capture failure, backed by focused tests and reproducible evidence.

After answering the four questions: is a PR needed at all to achieve this outcome?

Yes. The contributor confirmed this scope and requested submission after validation. It requires the shared execution change, root regressions and recipe updates.

For a bug fix: link actual observed failure evidence.

This is a capability extension. The new GPU test fails on main because resident execution makes no capture, then passes here. The exact red/green logs are in the qualification package below.

Purpose

CUA_S1_DEVICE_FEATURES=1 previously always used language eager even with CUA_S1_GRAPH=1. Route host/device staging through one existing multimodal capture/replay helper. Current IDs, T/H/W positions and image rows are staged into owned scratch before replay; external producer pointers remain outside captures. Scratch growth, synchronization, bounded-cache retirement and eager fallback are preserved. No kernel, ABI, response, scheduler or default change.

The latest update is documentation only: explain complete-working-set warmup, language scratch invalidation, vision entry/byte limits, and the actual scope of the measurements. No universal acceleration or default-on recommendation.

Test Plan

Base: 05075d0f36e399e3a001ffedea400c666990cda9. Current head: 911faa0d22753be56c8bf513a23741dfa6d00593. Runtime is unchanged from a14e1027e7de7507e6ff82b3844a916ff19a26b9.

• Current documentation update: format and strict MkDocs builds passed; full diff reviewed.

• Historical runtime qualification on RTX4090/SM8.9, driver595.71.05,CUDA13.0,Rust1.98,ABI9: Graph tests cover changed IDs/features/T-H-W positions, producer release, host/device interleave, scratch growth and recording failure; Cua device checks with visionGraph0/1; real queued-copy drain/release and recovery; frontend forwarding/failure spots. Pinned BF16 merged language and original BF16 vision/FP32 adapter; all artifact hashes verified.

• Original runtime local checks:360 workspace tests passed,0 failed,71 ignored, plus strict Clippy/release/docs/contract checks. These are historical checks of the runtime source, not newly executed GPU tests of the documentation commit.

Test Result

All normal responses match the eager control byte-for-byte. The original mixed matrix measured resident eager p50/p95=121.967/275.370ms, Graph124.894/276.195ms. Follow-up log alignment found5 timed language captures after scratch growth in its first round and15 timed vision captures(3/6/6). Those samples remain preserved and are not pure warmed replay.

The complete-warming follow-up excludes8 growth+8 all-shape stabilization+5 ordinary warmups, then measures3×8 requests on the same worker,concurrency1. Both entry and byte budgets must hold the working set: the1024-square image alone needs262537216 requested scratch bytes, so4 entries/256MiB churns this six-grid matrix. The explicitly configured8 entries/1024MiB fits it and removes timed captures; these are workload-specific controls, not changed defaults.

Same full-warming protocol, cache8/1024MiB	p50 ms	p95 ms	sampled process maximum MiB
resident eager	116.891	271.109	10138
resident Graph	116.399	269.011	10202

The ~0.4% p50 difference in this bounded sample does not establish a stable general win. Graph removes submission work; GEMMs, CPU preprocessing/readout, feature checks/staging and synchronization remain. The feature supplies an opt-in execution capability; defaults remain off. Clock=POST through complete JSON/schema validation; p95=pooled nearest rank; memory is sampled nvidia-smi process maximum, not allocator-exact peak.

Demo / evidence

Follow-up qualification: SHA256 e63be7ff5d2b105d9e19d9a1d88e8ee5283d45c7d46c1a9933d4cf258ca507da. Contains original-cache/full-cache raw and excluded follow-up HTTP samples, source pins and immutable source-archive links, cache/capture traces, protocol and reproduction controls. The unchanged a14e102 runtime was measured before the documentation commit; no fresh GPU run of this SHA is claimed. Expanded prefix numerical records in that package concern PR184 and are explicitly separate.

Initial qualification remains unchanged: original archive SHA256 933ee5749a2f1e02041f1087a992301e9206da44c7bf5a7ba59daf9db7db6e15, with runtime red/green/failure logs, source snapshots and offline audit. Original screenshot pixels/operator details are omitted from the new public package. Video:N/A. General GUI task quality and other hardware support are not established.

Self-review

Full diff and claims reviewed; no remaining actionable runtime finding. Current PR:110 authored Rust/test changed lines,169 total changed lines. The update changes31 documentation lines only; raw run payload remains external. Agent review does not replace the contributor's human attestation; keep draft and leave the checklist for that review.

• I have reviewed the full diff and addressed the issues I found.
• I have checked that the change follows the project's architecture and stays focused on the stated purpose.
• I have run the checks appropriate to this change and reported commands, results, and anything I could not verify above.
• I have checked that the PR description, documentation, and any accuracy or performance claims match the implementation and available evidence.