All pull requests

Pull requests

• feat(vulkan): dual-GPU 256K MoE serving prototype / 双卡 256K MoE 服务原型

  #106·a398289244-hue opened Oct 21, 2026·

  ·

• error: 20261020T075946.502Z-51ee5b8c83a55a250c33ff06ea164252

  #100·Headmaster218 opened Oct 20, 2026Owner·

  ·

• feat: register glm5next (GLM-5.3-Flash) architecture + CPU perf fixes

  #95·dingyiyuan opened Oct 16, 2026·

  ·

• feat(kv): paged frame-chain prefix cache (kv.paged_cache_max)

  #94·not-zen-girl opened Oct 16, 2026·

  ·

• [codex] feat(kv): RAM radix prefix cache for idle conversation KV

  #86·not-zen-girl opened Oct 15, 2026·

  ·
  2

• perf(mtp): calibrate the Qwen3.8 VERIFY width instead of trusting spec.k

  #72·not-zen-girl opened Oct 9, 2026·5 tasks done·

  ·

• feat(uncensor): Qwen3.8 refusal-direction projection with a per-request switch

  #67·not-zen-girl opened Oct 7, 2026·

  ·
  1

• Add qwen35/qwen35moe vision, fix server for image, add image formats

  #62·scriptnapps opened Oct 2, 2026Contributor·

  ·

• feat(paging): INFR_WARM_PAGE_CACHE to pre-read the shards at startup

  #61·HuangShanglong opened Oct 1, 2026Contributor·

  ·
  1

• perf(linux): stop aliasing the host tier through VK_EXT_external_memory_host

  #60·HuangShanglong opened Oct 1, 2026Contributor·

  ·
  1

• fix(hostmem): cap a host arena against physical RAM, not a 3 GiB tail

  #59·HuangShanglong opened Oct 1, 2026Contributor·

  ·

• feat(linux): add the Linux build guide, launcher and smoke test

  #58·HuangShanglong opened Oct 1, 2026Contributor·

  ·
  6

• Linux

  #44·c00jsw00 opened Sep 23, 2026·

  ·
  2