Add instance-local rule buckets to the Cua-S1 multimodal worker

[Levius-Fubuki]

Purpose

Add opt-in rule-bucket execution inside the model-owned Python worker. Nearby lengths can reuse a bucket after strict full-vocabulary-logit checks; per-length gates, resource ownership and shutdown remain instance-local. The latest follow-up integrates the graph lifetime fixes from #22/#33.

Stacked on #33. Incremental runtime comparison. The PR diff remains limited to runtime Python source; validation helpers, tests and generated evidence are preserved separately. This optimizes the Python multimodal worker, not the native Rust/CUDA worker.

Test Plan

System1-Omni Version / Commit: d598de3ce03f57dd20ff9b730279a4b4979623d4.

Fresh validation started 2026-09-30 on RTX 4090 24 GiB, driver 595.71.05, Torch 2.14.0+cu130, Transformers 5.17.0 and PEFT 0.21.0; BF16 base with unmerged adapter, without FLA/causal-conv1d.

• Validation checkout and pinned setup: PYTHONPATH=src:recipe/cua_s1 python -m pytest tests/cua_s1 -q.
• Exact GPU commands and exit statuses, using upstream-verified weights and a fresh output directory.
• Runtime/validation source mapping: git diff <runtime> <validation> -- src is empty. Source hashes and numerical/performance reports are independently verified.

Test Result

• CPU suite: 227 passed in 21.31s.
• Fresh full-checkpoint validation passes 136 full-vocabulary-logit comparisons and 20 complete responses, plus 255/256/257/319/320/321-token boundary, eviction and budget checks. Real HTTP changed-input parity, independent-model isolation and shutdown pass.
• The additional full-model eviction checker confirms three surviving-graph replays after actual eviction with changed contents and retained references.
• Two cold-cache runs across four hot lengths and twelve rotating lengths give 720 timed predictions, all with complete response equality. Bucket total time is 27.3–27.9% below eager on these samples. It is 1.5–1.9% below exact for hot lengths and about 25.0–25.1% below default exact for the rotation; rotation p95 is slightly above eager.
• Default exact makes no captures in the twelve-length rotation because the revisit interval exceeds its eight-request admission window. Add automatic Graph selection with shared Cua-S1 worker budgets #38 includes a controlled exact-window-32 comparison. These synthetic results support explicit opt-in, not general adoption or concurrency claims.
• GitHub rust CI passed on the above runtime head.

Full evidence and tradeoffs · Raw primary report · Independent verification summary.

Timing is serial and synthetic on one GPU. Cold-cache totals include capture and parity-check costs; model loading, HTTP parsing and final cleanup are excluded. Two runs are not confidence intervals or proof of production workload behavior.

Self-review

Agent-assisted full-diff review and validation limits found no actionable correctness or architecture defect. Runtime scope, commands and claims were checked against the recorded sources/results. The contributor remains responsible for understanding the changes; this does not replace maintainer review.

[hsliuustc0106]

Review of 3254e902eb86 (incremental stack changes).

Reviewed bucket padding/trimming, the pinned Transformers forward integration, per-real-length parity checks, cache growth and worker shutdown. No independent actionable correctness defect found in the inspected paths.

I would hold adoption until real GPU parity and controlled performance measurements establish the benefit of rule bucketing over the exact-length/eager paths, including varying lengths within a bucket and eviction/replay. This adds execution logic to the Python multimodal worker and does not accelerate #19's native Rust/CUDA worker.

Validation: 227 archived CPU tests passed against this head's runtime source using Transformers 5.17.0 and tokenizers 0.23.2. Six initial failures with Transformers 5.14.1 were version-guard failures and disappeared with the pinned version. Local Torch remains 2.13 rather than pinned 2.14. Tiny real model tensors were exercised on CPU; CUDA capture/stream behavior was mocked. No full-checkpoint GPU parity or new latency measurement was performed.

The validation suites were retrieved from the immutable archived revisions linked in the PR descriptions; they are not retained in the current PR diffs.