feat(cuda): add shareable cuMem IPC lifecycle

[devinkuhn]

Summary

Adds the direct CUDA-IPC/cuMem lifecycle required by LMCache multiprocess serving on the Jovian Judgement GLM-5.3 stack.

• POSIX-FD cuMem export/import with same-UID SCM_RIGHTS broker
• allocation mapping deduplication and alias refcounts
• retryable unregister/re-register without cudaDeviceReset
• partial-registration and partial-wrapper rollback
• active STORE/RETRIEVE leases so unregister/reaping cannot unmap live transfers
• stream-ordered failed-store reservation abort
• scheduler heartbeat lifecycle and bounded completion dedup
• failed registration preserves the prior live transfer context
• recurrent kernel groups receive independent object groups when separation is enabled
• padded physical block strides and retained-window retrieval from the prerequisite geometry commit

Dependency

This branch includes the already-submitted padded hybrid geometry prerequisite (1a31bb39). The new feature commits are:

• d016561a recurrent object-group isolation
• 1fee4175 shareable cuMem IPC lifecycle
• 40e28112 scheduler heartbeat lifecycle
• 05d3fb01 context lifetime hardening
• c43866fa transfer completion invariants

Runtime qualification

Four RTX PRO 6000 Blackwell GPUs, vLLM MTP3 TP4/DCP4, exact 1M context, batch 8192, FP8 KV:

• one LMCache server PID mapped on GPUs 0-3
• 4 physical imports / 184 aliases
• object groups [0,1,2,3]
• cold 32k coherent; 64 L1 objects / 818,151,424 bytes
• vLLM-only restart automatically unregisters to 0 imports/aliases while preserving L1
• warm reload coherent with 24,576 external cached tokens
• unrelated raw sentinel coherent
• 524,288-token request coherent at the qualified 31.25 GiB KV/GPU profile
• no Xid, illegal access, EngineDead, OOM, or restart in final accepted gates

Tests

• final focused regression set: 160 passed
• independent exact-image selected set: 116 passed
• prior cuMem/object-group suite: 129 passed
• Ruff format/check, py_compile, git diff --check passed

Duplicate-work note

No open LMCache PR implements this D22-style shareable cuMem broker/import lifecycle. Existing vLLM PR LMCache#526 carries the coupled vLLM/Jovian integration work.

AI assistance disclosure

AI assistance was used in preparing this contribution.

[coderabbitai]

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

🗂️ Base branches to auto review (1)

• dev/*

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Team

Run ID: dc4d07a9-d0db-4f91-928f-a2bd8dbcc02a

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}