Skip to content

fix(dcp): support disposable A2A channel profiling - #1

Draft
FujitsuPolycom wants to merge 1 commit into
masterfrom
codex/dcp-a2a-profile-lifecycle
Draft

fix(dcp): support disposable A2A channel profiling#1
FujitsuPolycom wants to merge 1 commit into
masterfrom
codex/dcp-a2a-profile-lifecycle

Conversation

@FujitsuPolycom

@FujitsuPolycom FujitsuPolycom commented Jul 20, 2026

Copy link
Copy Markdown
Owner

Summary

  • add opaque channel checkpoints to PCIeDCPA2APool
  • restore the exact pre-profile channel and alias map after disposable capture
  • close only channels created after the checkpoint
  • reject cross-pool checkpoints and lifecycle calls during active capture

This supplies the lifecycle contract used by vLLM's disposable DCP A2A memory
profiling without leaking profiling-only IPC allocations into the serving pool.

Sparkinfer rebase

The branch is based on local-inference-lab/sparkinfer@ec2cd4d and uses the
new sparkinfer.comm.pcie package layout. It does not restore the retired
b12x.distributed namespace.

Validation

  • focused lifecycle suite: 7 passed
  • combined Sparkinfer integration suite: 56 passed, 4 GPU-gated skips
  • Ruff and compileall: passed
  • included in the exact v19 integration image used for the TP4/DCP4 model gate

Fork integration draft only. Nothing has been pushed to the upstream repository.

AI-assisted implementation under human direction; the diff and validation
results were reviewed before publication.

Design document

GLM-5.2 Sparse CKV under Decode Context Parallelism is the canonical architecture, transport-contract, validation, and upstream-decomposition reference for this PR stack.

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.

1 participant