Skip to content

feat(kv): add Fluxon KV integration for SGLang - #42

Merged
Kaimary merged 33 commits into
mainfrom
sglang-fluxon-kv-integration
Oct 27, 2026
Merged

Kaimary merged 33 commits into
mainfrom
sglang-fluxon-kv-integration

Conversation

@zTz01

@zTz01 zTz01 commented Jul 8, 2026

Copy link
Copy Markdown
Collaborator

No description provided.

Align the bundled GPU communication SDK and consumer with schema 8 / ABI 10
so process logging and runtime requests use the same protocol. Include the
CUDA runtime libraries and remove local build paths from the SDK artifact.

Preserve GPU readiness retries by matching the typed transfer error. Install
libnvidia-compute-580 and check libcuda loading in all three CI jobs.

Validated SDK ABI checks, process logging, and external client CPU CRUD.
The GPU SDK also passed logging and CPU CRUD in an Ubuntu 24.04 container
with no NVIDIA devices and only the userspace driver dependencies.
zTz01 added 5 commits October 25, 2026 16:13
…acts

Route native writes through Owner allocation while retaining Master key
reservations, lease publication, overwrite ordering and shutdown cleanup.
Preserve the SGLang fast paths and enable ordinary CPU Owner configs.

Close asynchronous Get targets by exact generation when Delete invalidates
them, retain active holders, and bound GetDone admission and discovery.
Restore native Get retries for stale source plans without retrying completed
batch items or suppressing other errors.

Restore the upstream CPU storage rounds and update Owner reclaim coverage,
including SSD reload, stale routes, cancellation and concurrent overwrites.

Validation: 396 KV tests passed; memory, SSD and memory+SSD CPU integration
rounds passed. GPU/RDMA and full SGLang experiments were not rerun.
Pass create-only PUT options through Python and PyO3, including batch paths.
Retain the host framework for each P2P dispatch and reject callbacks after its release.
Wait for asynchronous holder release before checking GET/DELETE slot reclamation.
Prefer the linked SDK over wheel copies when running Cargo tests.
Use accepted Owner allocator reports for public and exported segment capacity. Preserve generation-fenced tomb cleanup.

Release transfer-engine dependencies before P2P and cluster handles, and retain each release operation until completion so close and canceled-close retries cannot leak IPC subscribers.

Synchronize the MQ capacity test with the asynchronous prefix index before probing backpressure.
Treat IPC PID records already removed by another startup cleaner as complete,
while preserving other I/O errors and live-process identity checks.

Rebuild the full GPU communication core from the same source snapshot; retain
ABI 10, schema 8, GPU/UVM behavior, and the existing native dependencies.
@Kaimary
Kaimary merged commit 5fe7226 into main Oct 27, 2026
8 of 14 checks passed

This branch had an error being deployed

1 failed deployment
OPENAI_API_KEY — b311cdd1 Deployed Oct 26, 2026 by ActivePeter via Codex failure analysis #516
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants