All pull requests

Pull requests

• fix(mamba): read the prefill donation guard from the CPU verify plan

  memory-pool
  #65·rchalamala opened Sep 26, 2026·

  ·
  1

• fix(mamba): skip prefill donation when an in-flight verify aliases the keep slot

  #64·rchalamala opened Sep 26, 2026·

  ·
  6

• feat(sampling): policy-driven deterministic top-p/top-k renorm

  #63·rchalamala opened Sep 26, 2026·

  ·
  18

• [Feature] Add TP-transaction hardening for MLA HiCache host dedup

  hicacheunified-radix-cache
  #62·rchalamala opened Sep 26, 2026·

  ·
  19

• [Feature] Add opt-in MLA/DSA HiCache host dedup across attention-TP ranks

  hicacheunified-radix-cache
  #61·rchalamala opened Sep 25, 2026·

  ·
  24

• [Fix] Preserve full multimodal content identities across caches

  dependenciesdocumentationhicacheMulti-modalunified-radix-cache
  #55·rchalamala opened Sep 25, 2026·7/7·

  ·
  14

• [Fix] Preserve multimodal transport and session ownership

  #54·rchalamala opened Sep 25, 2026·6/7·

  ·
  8

• [Fix] Balance CUDA IPC export ownership

  #53·rchalamala opened Sep 25, 2026·5/7·

  ·
  11

• [Fix] Contain invalid target verification in speculative decoding

  hicachejit-kernelunified-radix-cache
  #52·rchalamala opened Sep 25, 2026·2/2·

  ·
  11

• [Fix] Synchronize multimodal cache misses across attention ranks

  #51·rchalamala opened Sep 25, 2026·4/7·

  ·
  5

• [Fix] Preserve complete DFlash penalty history across overlap forwards

  #50·rchalamala opened Sep 25, 2026·

  ·
  9

• [Fix] Stop OpenAI engine-fault responses without final usage

  #49·rchalamala opened Sep 25, 2026·2/2·

  ·
  7

• [Fix] Join HiCache backup completion before subsequent forwards

  hicache
  #48·rchalamala opened Sep 25, 2026·

  ·
  1

• [Perf] Cache Kimi chat segments within a bounded budget

  #47·rchalamala opened Sep 25, 2026·

  ·
  5

• [Perf] Optimize short Kimi tokenizer segments

  #46·rchalamala opened Sep 25, 2026·

  ·
  7

• [Fix] Handle truncated Kimi K3 opening-think suffixes

  #45·rchalamala opened Sep 25, 2026·

  ·
  4

• [Feature] Record host gaps between decode steps

  #44·rchalamala opened Sep 25, 2026·

  ·
  15

• [Fix] Enforce exact multimodal embedding rows at final placement

  #43·rchalamala opened Sep 25, 2026·

  ·
  1

• [Fix] Validate caller token IDs before multimodal expansion

  npu
  #42·rchalamala opened Sep 25, 2026·

  ·
  11

• [Feature] Expose scheduler generation timestamps in native metadata

  #41·rchalamala opened Sep 25, 2026·

  ·
  5

• [Feature] Record Mamba checkpoint gaps at first prefill admission

  unified-radix-cache
  #40·rchalamala opened Sep 25, 2026·

  ·
  7

• [Feature] Add opt-in sampled FP8 query range observations

  blackwell
  #39·rchalamala opened Sep 25, 2026·

  ·
  12

• [Feature] Add normalized finish fields to request metrics records

  #38·rchalamala opened Sep 25, 2026·

  ·
  1

• [Feature] Add logical host pool occupancy metrics including sidecars

  hicache
  #37·rchalamala opened Sep 25, 2026·

  ·
  11

• [Feature] Expose effective scheduler capacity metrics

  #36·rchalamala opened Sep 25, 2026·

  ·
  7