[PR #5] [MERGED] [Bugfix] Prevent duplicate Mooncake push metadata across scheduler steps #1

New issue

Closed

opened 2026-10-29 21:06:01 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-29 21:06:01 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/kvcache-ai/vllm/pull/5
Author: @jiaran-king
Created: 8/26/2026
Status: ✅ Merged
Merged: 8/27/2026
Merged by: @gty111

Base: add-mooncake-ec-connector ← Head: fix/mooncake-push-metadata-dedup

📝 Commits (1)

456f1f2 [Bugfix] Prevent duplicate Mooncake push metadata across scheduler steps

📊 Changes

2 files changed (+28 additions, -1 deletions)

View changed files

📝 tests/v1/ec_connector/unit/test_ec_mooncake_connector.py (+12 -0)
📝 vllm/distributed/ec_transfer/ec_connector/mooncake_ec_connector.py (+16 -1)

📄 Description

Purpose

The producer scheduler currently deduplicates a push transfer only within one
scheduler step. build_connector_meta clears _pushes_to_prepare after emitting
the push specs, so revisiting the same request on a later scheduler step can
queue the same transfer_id again.

The repeated metadata reaches start_save_caches, causing redundant reservation
work for the same transfer.

This change:

tracks prepared transfer IDs across scheduler steps;
emits the initial push metadata once per request lifetime;
removes the tracking entries in request_finished; and
leaves the existing worker-side reservation refresh unchanged.

No worker/consumer reservation semantics, lease settings, control protocol,
cancellation, receive-pool, or retry behavior are changed.

Duplicate-work Check

I searched open vLLM PRs for mooncake duplicate push metadata,
mooncake transfer_id scheduler, and ECMooncakeConnector duplicate. No open
PR addresses this request-lifetime deduplication bug. This change targets the
behavior currently present in the source branch of #41567.

AI assistance was used in preparing this change.

Test Plan

.venv/bin/python -m pytest \
  tests/v1/ec_connector/unit/test_ec_mooncake_connector.py \
  -q

The regression test verifies:

the first scheduler step emits the push spec;
revisiting the same request on a later step does not emit it again; and
request_finished removes the tracking state.

Test Result

42 passed, 1 warning

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/kvcache-ai/vllm/pull/5 **Author:** [@jiaran-king](https://github.com/jiaran-king) **Created:** 8/26/2026 **Status:** ✅ Merged **Merged:** 8/27/2026 **Merged by:** [@gty111](https://github.com/gty111) **Base:** `add-mooncake-ec-connector` ← **Head:** `fix/mooncake-push-metadata-dedup` --- ### 📝 Commits (1) - [`456f1f2`](https://github.com/kvcache-ai/vllm/commit/456f1f28ae64ec4fb57060b0cba50dc621873883) [Bugfix] Prevent duplicate Mooncake push metadata across scheduler steps ### 📊 Changes **2 files changed** (+28 additions, -1 deletions) <details> <summary>View changed files</summary> 📝 `tests/v1/ec_connector/unit/test_ec_mooncake_connector.py` (+12 -0) 📝 `vllm/distributed/ec_transfer/ec_connector/mooncake_ec_connector.py` (+16 -1) </details> ### 📄 Description ## Purpose The producer scheduler currently deduplicates a push transfer only within one scheduler step. build_connector_meta clears _pushes_to_prepare after emitting the push specs, so revisiting the same request on a later scheduler step can queue the same transfer_id again. The repeated metadata reaches start_save_caches, causing redundant reservation work for the same transfer. This change: - tracks prepared transfer IDs across scheduler steps; - emits the initial push metadata once per request lifetime; - removes the tracking entries in request_finished; and - leaves the existing worker-side reservation refresh unchanged. No worker/consumer reservation semantics, lease settings, control protocol, cancellation, receive-pool, or retry behavior are changed. ## Duplicate-work Check I searched open vLLM PRs for `mooncake duplicate push metadata`, `mooncake transfer_id scheduler`, and `ECMooncakeConnector duplicate`. No open PR addresses this request-lifetime deduplication bug. This change targets the behavior currently present in the source branch of #41567. AI assistance was used in preparing this change. ## Test Plan .venv/bin/python -m pytest \ tests/v1/ec_connector/unit/test_ec_mooncake_connector.py \ -q The regression test verifies: 1. the first scheduler step emits the push spec; 2. revisiting the same request on a later step does not emit it again; and 3. request_finished removes the tracking state. ## Test Result 42 passed, 1 warning --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-29 21:06:01 +00:00