Reuse Cua-S1 image preprocessing and vision features per request

[Levius-Fubuki]

Purpose

A request with eight questions over one screenshot currently preprocesses and encodes that image eight times. This change performs image preprocessing and adapted vision encoding once per multi-question request, while preserving independent prompts, candidate order, text embeddings, 3D positions and language forwards. Single-question requests retain the original path. Every prompt is validated before inference; features remain request-local, including failure paths.

On RTX 4090, 17 paired workloads / 3,400 timed requests show 6.8–19.9% lower p50 latency for multi-question cases. Eight distinct questions improve from 1100.03 / 1094.87 ms to 887.54 / 888.45 ms across two runs. Single-question p50 remains within measured noise (−0.5% to +1.1% reduction).

Depends on #12 and #15. This branch includes their commits until they merge; this PR's new work begins after 4c605e3. It is a Transformers/PEFT request-local reuse optimization, not a native CUDA backend or custom kernel.

Full results, raw samples, tensor interface and reproduction commands.

Test Plan

• Compare exact prepared tensors, language embeddings, attention masks, 3D positions and complete responses against the original execution path. Include distinct questions, candidate permutations, 1–26 candidates, PNG/JPEG, Unicode and consecutive image changes.
• Alternate baseline/reuse sample order; synchronize each prediction; keep instrumentation outside timed calls; record per-variant memory peaks, source and environment.
• Rerun the pinned independent upstream oracle, test the actual HTTP worker, and independently recompute sample counts and percentiles from saved JSON.
• Run the full Cua-S1 test suite, Ruff lint/format checks and failure-recovery tests. Stabilize the existing header-rejection test's body-write race without changing server code.

System1-Omni Version / Commit: timed clean source c18a21b; final GPU-host CPU/oracle/HTTP checks at clean 9f4a4fe. Later commits add result evidence and documentation; measured model and benchmark code are unchanged.

Test Result

• 104 tests passed on the GPU host; 103 passed, 1 skipped locally (torch absent locally; that tensor test passes remotely). Lint/format checks pass.
• 13 correctness fixtures / 49 questions per path, plus validation of all 17 timed cases: exact input, embedding, position and full-response parity. Multi-question preprocessing/vision counts are N → 1; language counts remain N.
• All 9 fresh independent upstream forwards match exactly. Actual HTTP health and eight-question inference return 200 with exact engine-response parity.
• 3,400 timed requests / 13,600 language forwards; all counts, alternating orders, p50/p95 values and reference matches independently verified. Largest allocated GPU peak is 9.083 GiB baseline / 9.035 GiB reuse.
• Complete evidence archive copied off the GPU host and SHA-256 verified. Results are synthetic, concurrency-1 engine latency; HTTP throughput, p99, native kernels and Metal are outside this claim.

[Copilot]

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