[None][feat] RI-02 bind shared KV leases to lifecycle retirement

[chienchunhung]

Summary

Bind shared-cache transfers to manager-owned staging leases and existing lifecycle retirement. A failure, cancellation, request exit, or timeout can become visible immediately while the adapter retains memory until backend access and local copies are proven complete.

Scope

• Add SharedStagingAdapter, rooting the manager, lender, PartsHold, registrations, ready lease, extent and Attempt before backend exposure.
• Reuse LC-MC-0 RetirementDeadline arbitration, poll logical outcomes independently, and collect potentially blocking backend quiescence off the owner thread. Rejected submissions and ambiguous escaped submissions have distinct cleanup behavior.
• Mark only served whole rows after positive backend quiescence, then record a CUDA event on the exact manager stream; its completion proof survives request exit. Shutdown drains access/copies before deregistration and PartsHold release.
• Add CPU lifecycle tests and nine real-manager GPU cases covering complete/partial/missed delivery, failure/cancellation, delayed copies, request exit/page reuse and source retention. Document the first-profile restriction: one ready operation per deadline and one registered provider per adapter; no scheduler activation or new lifecycle state machine.

Verification

At fe52dac001557bbf8a764b69a6bcaa955bb17a68:

• Complete selected normal-runtime suite: 136 passed, 0 failed, 0 skipped in 12.42s on one B300 PCIe (ComputeLab allocation 4681685), Python 3.12 / CUDA 13.4 / Torch 2.14 NV26.08. All 9 real-manager GPU cases executed; pytest thread-leak checking remained enabled.
• Coverage: backend contracts 25, staging mapping 42, resource lifetime 31, lifecycle binding 10, lender naming 19, GPU adapter 9.
• Built and installed TensorRT-LLM 1.4.0rc0 from 7673bb837a3c2e5e8457f158de72ca2e24fc710a; the final amendment changes only test cleanup. Git-tree equality and loaded module/native-binding hashes verified compatibility with the final head.
• Earlier run: 133 passed / 3 failed because three CPU tests returned with active proof workers. They now finish and join those workers before returning; all assertions remain enabled.
• Local focused CPU suite: 117 passed. Installed commit hooks and git diff --check: passed. Fresh complete self-review after the last fix: no actionable findings.
• Full CI launched for this head: PR_Github #76146, with bot acknowledgment. The pipeline is pending; no passing result is claimed.
• These module tests do not qualify a real NIXL provider or enable the runtime profile; LC-MC-21 / RI-09 remain separate.

Notes

Stacked development prerequisites: #19722 lender → #19823 RI-01; #19822 LC-MC-0 is also required. This draft targets main and includes those separate commits. Review RI-02-only commit for its four-file scope; rebase after prerequisites merge.

• Reviewed repository PR checklist for this scope; tests and developer documentation included. No protected user-facing API or new dependency changes.

[chienchunhung]

/bot run

[tensorrt-cicd]

PR_Github #76146 [ run ] triggered by Bot. Commit: fe52dac Link to invocation

[tensorrt-cicd]

PR_Github #76146 [ run ] completed with state FAILURE. Commit: fe52dac
/LLM/main/L0_MergeRequest_PR pipeline #62775 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation