All pull requests

Pull requests

• Serve Kimi K3 LoRA adapters on MXFP4 Marlin experts

  #97·micahtyong opened Nov 12, 2026·

  ·

• [Feat] beam: support LoRA on the Spindle v0.5.23 fork

  lorarun-ci
  #95·frrad opened Nov 10, 2026·3 of 5 tasks·

  ·
  1

• Remove Spindle scheduler snapshot hook from v0.5.23

  #94·kailash109 opened Nov 10, 2026·

  ·

• Remove Spindle's scheduler snapshot hook in favor of native KV events

  #93·kailash109 opened Nov 10, 2026·

  ·

• Restore MXFP8 and BF16 MoE checkpoint layouts before reloads

  quant
  #91·kailash109 opened Nov 9, 2026·

  ·

• [Feat] [NVIDIA] megakernels: Expose the KDA verify ring and decode-side MLA metadata to plugins

  #88·vineelbhat opened Oct 29, 2026·2 of 5 tasks·

  ·

• Apply CSGMV split-K shrink to every LoRA shape, scaled by input size

  jit-kernel
  #87·kailash109 opened Oct 28, 2026·

  ·
  1

• [Fix] spindle-routing: Keep routing observations from stopping the scheduler

  #86·kailash109 opened Oct 26, 2026·

  ·
  4

• [Fix] constrained: Add backend-level vocab-mask hooks to GuidanceBackend

  run-ci
  #83·clairec03 opened Oct 22, 2026·

  ·
  5

• Delegate LoRA hooks from the Nemotron-H VL wrapper to its language model

  lora
  #78·micahtyong opened Oct 15, 2026·

  ·
  1

• Backport expert LoRA support onto stitch-sglang-v0.5.21

  lora
  #76·kailash109 opened Oct 13, 2026·

  ·

• [LoRA] Release adapter references exactly once per request state

  lora
  #75·micahtyong opened Oct 9, 2026·

  ·

• Add opt-in error for greedy fallback in speculative verification

  #74·jvmncs opened Oct 6, 2026·3 of 5 tasks·

  ·

• [Kimi K3] Stream tool-call arguments; DFlash selector + penalizer review follow-ups

  jit-kernelrun-ci
  #72·devin-ai-integration[bot] opened Oct 6, 2026Bot·3 of 5 tasks·

  ·
  7

• [SMG] Clear cache-aware trees for workers flushed by /flush_cache

  model-gateway
  #71·aryatsc opened Oct 6, 2026·2/2·3 tasks done·

  ·

• [SMG] Backlog-aware prefill selection for cache_aware in PD mode

  documentationmodel-gateway
  #70·aryatsc opened Oct 6, 2026·1/2·5 tasks done·

  ·

• Support exact DFlash rollout distribution capture for score centering

  jit-kernel
  #69·jvmncs opened Oct 3, 2026·

  ·
  1

• Allow tool choice auto and response format for GLM series

  #68·akelch11 opened Oct 3, 2026·5 tasks·

  ·

• Make GLM-5.3 Flash DCP production-ready

  amdblackwelldeepseekdocumentationhicachejit-kernelmemory-poolrun-cirun-ci-extraspeculative-decodingunified-radix-cache
  #66·willhu-jpg opened Sep 28, 2026·

  ·
  1

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

  memory-pool
  #65·rchalamala opened Sep 26, 2026·4 tasks done·

  ·
  1

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

  #64·rchalamala opened Sep 26, 2026·4 tasks done·

  ·
  6

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

  #63·rchalamala opened Sep 26, 2026·5 tasks done·

  ·
  18

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

  hicacheunified-radix-cache
  #62·rchalamala opened Sep 26, 2026·4 tasks done·

  ·
  19

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

  hicacheunified-radix-cache
  #61·rchalamala opened Sep 25, 2026·4 of 6 tasks·

  ·
  24

• [Fix] Preserve full multimodal content identities across caches

  dependenciesdocumentationhicacheMulti-modalunified-radix-cache
  #55·rchalamala opened Sep 25, 2026·7/7·3 of 4 tasks·

  ·
  14