Skip to content

dfkv v2.26.4

Choose a tag to compare

View all tags
@github-actions github-actions released this 11 Sep 20:33
· 28 commits to main since this release
v2.26.4
34ca8fd

SGLang DeepSeek-V4.1 HiCache deployment guidance

Documentation and release-metadata update following actual SGLang + HiCache + dfkv cold/L3 acceptance. No connector runtime or functional C++ changes; the published HiCache plugin is byte-identical to the tested v2.26.3 plugin.

Corrected deployment guidance

  • Pin the supported SGLang preview image and use its model-selected attention/MoE backends.
  • Configure MAX_BLOCK_BYTES separately from MAX_PAYLOAD_BYTES. The default 4 MiB object declaration can reject a valid larger SWA object despite a larger payload ceiling.
  • Explicitly select the reachable RDMA data fabric. MDS discovery, MR registration and transport=rdma are not proof of successful data I/O.
  • Use ratio-based host allocation for the model's mixed pools; correct the preview engine's fixed-size unit to decimal GB.
  • Distinguish GPU/host write-through from asynchronous L3 backup and local cache salt from external-storage key isolation.
  • Describe side-pool completeness, DSpark layout isolation and measured, bounded RDMA depth tuning without changing defaults.

Verification scope

  • Two input-length classes and two concurrency levels, three repetitions per cell, repeated for two RDMA depth settings.
  • Separate cookbook DSpark low-latency comparison and fresh-process L3 restore.
  • 60 measured paired phases, 1,920 requests with fixed 256-token outputs; all cold phases required PUT and zero GET, all hot phases required native GET hits plus storage-tier token hits. Measured hot device/host cache tokens were zero; native GET misses and RDMA completion-error deltas were zero.
  • Separate exact-answer cold/hot/restart quality checks passed.
  • L3 benefits are workload-dependent: short single-request inputs did not improve; longer prefixes benefited. Server-side L3 used both RAM and backing storage, not a forced cold-SSD-only test.
  • PR/final-main CI and Release workflow passed; all five payload assets verified against SHA256SUMS. Published server/MDS/CLI report 2.26.4, with no dynamic libstdc++ dependency for MDS and server GLIBC requirement at most 2.34.

PR: #380
Full changelog: v2.26.3...v2.26.4

Assets 8
Loading