Skip to content

Support FlexKV Layerwise#428

Draft
Clebrate wants to merge 5 commits into
NVIDIA:mainfrom
Clebrate:lyl_FlexKV_layerwise
Draft

Support FlexKV Layerwise#428
Clebrate wants to merge 5 commits into
NVIDIA:mainfrom
Clebrate:lyl_FlexKV_layerwise

Conversation

@Clebrate

Copy link
Copy Markdown
Collaborator

Description

Add FlexKV layerwise transfer support for the RecSys KVcache manager.

This PR wires the RecSys FlexKV backend to FlexKV's layerwise launch path by:

  • adding an optional RECSYS_FLEXKV_ENABLE_LAYERWISE switch;
  • passing layerwise_transfer and counter_id to FlexKV onboard launches when enabled;
  • adding a RecSys-side eventfd sender for FlexKV layerwise synchronization;
  • covering the layerwise onboard path in test_flexkv.py.

Checklist

  • I am familiar with the Contributing Guidelines.
  • New or existing tests cover these changes.
  • The documentation is up to date with these changes.

@Clebrate
Clebrate force-pushed the lyl_FlexKV_layerwise branch from b2ae455 to 42ddf58 Compare June 23, 2026 01:47
@Clebrate
Clebrate force-pushed the lyl_FlexKV_layerwise branch from d75f83b to 9232b2d Compare July 7, 2026 13:07
Clebrate added a commit to Clebrate/recsys-examples that referenced this pull request Jul 8, 2026
Keep non-layerwise benchmark cleanup improvements (FlexKV client shutdown
and scenario1 offload drain) while moving layerwise benchmark and inference
integration to PR NVIDIA#428.
@Clebrate
Clebrate force-pushed the lyl_FlexKV_layerwise branch 8 times, most recently from b614e68 to 64f2c59 Compare July 10, 2026 07:47
shijieliu pushed a commit that referenced this pull request Jul 10, 2026
* Recsys-FlexKV three scenarios inference benchmark

* benchmark and ablation

* Remove layerwise inference benchmark from PR429.

Keep non-layerwise benchmark cleanup improvements (FlexKV client shutdown
and scenario1 offload drain) while moving layerwise benchmark and inference
integration to PR #428.

* Consolidate FlexKV onboard benchmark mode.

Use --only-onboard to select the onboard-only benchmark path and add configurable request batch size support while removing the duplicate benchmark entry point.

* fix
@Clebrate
Clebrate force-pushed the lyl_FlexKV_layerwise branch from 64f2c59 to cdf053d Compare July 13, 2026 06:34
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