cua_s1: add native multimodal language CUDA Graph replay

[Levius-Fubuki]

Purpose

Enable opt-in native multimodal language CUDA Graph replay with CUA_S1_GRAPH=1. Token embeddings, BF16 image rows and T/H/W rotary tables are uploaded before replay. Vision remains eager. Each mode has a separate 64-entry FIFO; misses retain the eager result, and growth/drop or capture failure clears both caches. Failure diagnostics remain unconditional and capture failure disables Graph for both modes.

Rebased onto main 99865743d27316fbe81362dc0f0e6de6fda86284 to resolve the conflict with #102. Preserve its packed text prefill, ordered sequence-length cache keys, per-sequence math and output order. Text keys such as [4], [1,3] and [3,1] stay distinct; multimodal uses singleton [t] in its separate cache. The full upstream run() body remains unchanged.

The core diff still contains only src/models/qwen3_5/native/src/model.rs (63 insertions, 16 deletions). Root regressions and their wiring remain in dependent Draft #106. No kernel, ABI, dependency, weight, generated evidence or test body is added here. CUA_S1_GRAPH_TRACE=1 optionally logs successful capture/replay.

Test Plan

System1-Omni Version / Commit: base 99865743d27316fbe81362dc0f0e6de6fda86284; core head 3ef72e4d79ce2189532a8da2373bfb2d0937273e; dependent test head 2dd88db234b99a64061dbfd0094a4ab6b0b4e5cb.

cargo fmt --all --check
cargo clippy --workspace --locked --offline --all-targets -- -D warnings
cargo test --workspace --locked --offline
cargo build --workspace --release --locked --offline

Updated root GPU tests and setup. The runner executes three ignored cases serially with tracing unset and requires both unconditional failure diagnostics. Coverage includes current token/image/position uploads, misses/replay, text/multimodal interleave, scratch growth, FIFO eviction/recapture and recording-error recovery. #106 now also checks packed text [4] / [1,3] / [3,1], changed IDs and same-shape eager/replay outputs without aliasing equal total lengths. This does not claim bitwise batched/independent eager equivalence or cover every CUDA failure class.

Test Result

Fresh local checks above all passed for the rebased core: 90 passed, 0 failed, 13 ignored. The updated dependent tests also pass all four checks: 90 passed, 0 failed, 16 ignored; the three new GPU cases compile/discover but are ignored in normal CPU execution. Complete rebase/self-review logs, source hashes, patches and preserved old-test compilation failure. The core-prefix and unchanged-upstream-run byte checks pass.

Current-head GPU execution, performance and recording are unverified. The previously supplied GPU server refused SSH connections; no paid restart or new campaign was performed. Prior head b418230f and test head 975201ca passed three explicit GPU regressions and both diagnostic gates; those results are historical and do not validate the expanded cases/current rebased head. New-head GitHub Rust, benchmark and docs CI all passed (deployment skipped). Old-head CI is not carried forward as current validation.

Demo / evidence

Historical recordings, complete raw results, model/adapter pins, configuration, licenses and provenance are retained at their original revisions. The latest two English 30-second videos were recorded from b418230f after warm capture, with genuine HTTP choices/clicks, no added waits and serial panels shown at 1x. They were not re-recorded for this conflict repair.

On that historical head, the isolated same-native Graph ON/OFF/ON workload (90 frozen requests x2, 10 excluded warmups, 1024 tokens, concurrency1, RTX4090/BF16; vision eager) measured 133.1753 /144.3275 /133.9943 ms medians: 7.44% lower warmed HTTP latency, with all 360 ON/OFF response pairs byte-identical. Separate whole-framework native/reference comparisons were 2.48x text /1.51x multimodal and included earlier kernel/merged-LoRA/readout optimizations. These figures are not new-head performance claims. Capture cost, peak memory, universal speedup and agent accuracy were not established.

Self-review

Full repaired diff, #102 integration, architecture boundaries, cache lifetime/failure handling, changed test cases, artifact scope, source binding and available checks were reviewed again on 2026-10-07. No actionable source-level finding remains; unavailable GPU validation is disclosed above. The contributor explicitly confirmed these checklist items on 2026-10-06; that confirmation and Ready authorization are retained, with the new-head verification limits reported here.

• 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.

[hsliuustc0106]

审查了 head b418230f53a158fcbea706c1815c529a8b3832ec（base 47eff9cdeda01e4847a4fb9634a43f2cab6a233f），未发现可确认的代码缺陷。

• 输入在 replay 前重新上传；cache miss 保留 eager 结果，避免重复推进残差。文本/多模态缓存分离，扩容和捕获失败时的清理逻辑正确，视觉仍走 eager。
• 原始证据的源码 hash 与 PR 一致。重算确认 Graph ON/OFF 延迟下降 7.443%，360 组配对响应字节完全一致；该结果限于已预热、1024 tokens 的测试配置。
• 新录制日志显示文本 254 次、多模态 105 次调用均为 replay，没有新增 capture。证据包

一个维护建议：把独立 evidence 分支上的回归测试及注册另行合入主仓库，否则当前 CI 无法持续覆盖新增 graph 路径。

本地 cargo fmt --all --check、严格 workspace Clippy、workspace tests 均通过：89 passed、0 failed、12 ignored。Clippy 和 tests 使用 --locked --offline。GPU 测试核对了作者日志，未现场重跑；视频未做逐帧视觉检查。

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[hsliuustc0106]

fix conflicts please

[Levius-Fubuki]

fixed