laya: bound native workspace cache before allocation

[Levius-Fubuki]

Purpose

Configure retained limits through CacheConfig and Model::load_with_cache, preserving four-shape/512 MiB defaults. Estimate the allocator’s same 17-buffer layout and evict LRU workspaces before allocating a miss. Admit exact fits; zero limits or oversized shapes reuse one eager workspace. Model::clear_cache releases scratch while retaining loaded weights. Budget excludes weights, graph/driver overhead, host staging and eager fallback storage.

Stack 4/4, dependent on #73; merge in order. This PR targets main to run current-head CI and includes predecessor increments until they merge. Review this increment against 231cf358770e1464efbbce86ec6ce47cb5746056.

Test Plan

Published head: ae8e34fcd292d4c8f21c35d4d143033a8f56a796
Validated source revision: 307380eef58133851c614f0a5e0f7dc20041c070 (source-identical tree; the published head adds only an empty CI-trigger commit)
Incremental baseline: 231cf358770e1464efbbce86ec6ce47cb5746056
Current-main baseline: 7f39ac40902c374803992407bb26eeba29c8a588

cargo fmt --all --check
cargo clippy --workspace --locked --all-targets --offline --features omni-laya/serve -- -D warnings
cargo test --workspace --locked --offline --features omni-laya/serve
cargo build --workspace --release --locked --offline --features omni-laya/serve

Review incremental and cumulative diffs. Existing history is preserved: the previous PR head remains an ancestor. Current full-request processing/heads/HTTP/runtime integration replaces obsolete encoder-only modules.

Test Result

All four checks passed on the exact individual validated source revision above. The published head has the same tree, confirmed with Git; its current-head CI is checked separately. 83 passed, 0 failed, 10 ignored. Ignored checkpoint/GPU tests remain unverified. Incremental git diff --check passed. The PR remains core source only; external fixtures/logs/evidence are retained locally.

Seven external model fixtures passed all 160 supported-shape size estimates; exact/count/byte limits; allocation-peak verification proving preallocation eviction; LRU and validation before mutation; disabled/oversized eager reuse; clear/release/recapture and failed-capture admission cleanup/retry; unchanged launch trace and A→B→A staging updates. The complete backend equals real-driver-validated #73 byte-for-byte.

Fresh generic CUDA validation passed using exact #73 backend source 231cf358770e1464efbbce86ec6ce47cb5746056 on RTX 4090 (sm89), driver 580.105.08, nvcc 13.0.88 and Rust 1.99.0. An external init-only sm89 bootstrap and scalar-add kernel exercised production resource/copy/capture/replay/free operations. The production Hopper gate was preserved. #71 resolved dispatch matches tested methods byte-for-byte; #72 group copy matches outside the later capture guard; #74 backend source equals #73. Two setup attempts stopped before GPU work because Python was missing from PATH; corrected PATH passed with unchanged source/assertions.

Validation limit: This is generic host/backend FFI evidence, not Laya support on RTX 4090. Full optimized Laya eager/Graph/scorer/action-head/HTTP numerical parity and performance were not rerun on Hopper after this port. No new Laya accuracy or speedup claim is made. The earlier 2026-10-04 H800 encoder-stack comparison remains historical.

Numerical CUDA/TileLang/RoPE implementation, operator export/build sources, weights/conversion, preprocessing, padding, decoding, executor/serving and Cargo metadata match current main. Only host execution/resources/cache and scoped C capture cleanup change.

Self-review

Incremental and cumulative diffs were reviewed against current architecture/runtime contracts, supported inputs, precision, pointer/drop lifetimes, failure paths and claims. Local review found no remaining actionable defect in this host scope. Fresh full-Hopper validation remains unverified; contributor and maintainer review are separate from this assistance.

Demo / evidence

Raw current-source workspace, CPU trace/copy/capture/cache fixture logs, exact source/ancestry hashes and real-driver output are retained under artifacts/all-pr-followups-20261005/laya/, outside this core-code diff. The deliberately caught panic in the driver log is an expected cleanup test; the suite exit code is zero. No video or measured speedup is claimed. Current-head GitHub Rust, benchmark and docs CI all passed; deploy skipped. CI run.

Rebuild liblaya_cuda.so and regenerate its build-manifest hash to include the C capture cleanup change. The C ABI signatures remain compatible; the old library does not gain the C-side fix merely by rebuilding the Rust worker.

[Copilot]

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