Skip to content

v0.23.5

Latest

Choose a tag to compare

@xiaguan xiaguan released this 23 Jul 07:24
7a75ee6

Highlights

  • Cache pressure now prefers reclaiming remote-fetched KV replicas before local/SSD residents, reducing thrashing of primary copies under multi-node load (#412).
  • MetaServer insert responses can mark a newly registered replica as reclaimable once a block has three or more live owners, so redundant third-or-later copies are demoted for early eviction (#413).
  • Prefill/decode handoff can complete remote prefetch with partial tail blocks: decode loads prefill-saved partial prompt blocks while recomputing the logits token, and pegaflow.wait_for_full_prefix keeps the query in prefetch until the remaining prefix is ready (#411).
  • Added a NIXL P/D deployment example for GLM-5.2 FP8 with MTP and MultiConnector (#410).

Cache replacement policy

Source-based classes (from #412), refined by MetaServer hints (from #413):

Event Class behavior
Local save / SSD prefetch creates a resident block retained
RDMA/P2P fetch creates a resident block reclaimable
MetaServer insert: new owner and post-insert count ≥ 3 demote matching retained resident to reclaimable
Duplicate insert or cache hit refresh current class LRU; do not migrate
Memory pressure evict reclaimable LRU, then retained LRU

reclaimable_hashes is an insert-time signal, not a continuously synchronized replica count. Pega server and MetaServer are expected at the same version for this release.

Prefetch and P/D (#411)

  • pegaflow.wait_for_full_prefix keeps a query in PegaFlow's prefetch state until the remaining prefix is complete.
  • Decode can load prefill-saved partial prompt blocks and recompute the logits token.
  • CUDA IPC registration selects and synchronizes the correct device.
  • Documented P/D hash seed, hash algorithm, and connector settings.
  • Validated on H200 x8 with GLM-5.2 FP8: multi-turn and long-prompt runs showed partial-tail hits and greedy token parity with cold baseline (no TTFT improvement claimed).

Docs

  • NIXL P/D deployment example with shared hash/layout invariants and D-to-P reuse validation (#410).

Full Changelog: v0.23.4...v0.23.5