All pull requests

Pull requests

• fix: align SGLang cache lifecycle, eviction, and KV events

  #583·PeaBrane opened Oct 24, 2026Contributor·

  ·
  4

• chore: clean up stale B60 measurement data

  #579·mandy-li opened Oct 23, 2026Contributor·

  ·
  8

• feat(replay): add neutral hooks for native routing compositions

  #578·Harrilee opened Oct 23, 2026Contributor·

  ·
  3

• fix: correct the hybrid attention layer types of MiMo-V2-Flash and Llama 4

  #577·Arsene12358 opened Oct 23, 2026Contributor·

  ·
  3

• feat: add output adapter lifecycle callbacks

  #576·sttts opened Oct 23, 2026Contributor·

  ·
  5

• test: restore the root logger after each test

  #573·Arsene12358 opened Oct 23, 2026Contributor·

  ·
  3

• fix: support Qwen3.8-27B and Qwen3.8-27B-FP8 by correcting quantization

  #572·jfostier opened Oct 23, 2026·

  ·
  12

• fix: align SGLang admission and LRU boundary behavior

  #571·xuhuan51 opened Oct 23, 2026·

  ·
  17

• fix: adapt vLLM attention metadata to removed CPU fields

  #569·simone-chen opened Oct 22, 2026Contributor·

  ·
  1

• fix: qualify TRT-LLM rc29 collectors for Dynamo 1.6

  #568·simone-chen opened Oct 22, 2026Contributor·

  ·
  1

• fix: align learned FPM defaults with gym recipes

  #567·diffractiongoo opened Oct 22, 2026Contributor·

  ·
  19

• fix: defer unknown cache layouts to the runtime probe in onboarding

  #565·Arsene12358 opened Oct 22, 2026Contributor·

  ·
  12

• fix: release agentic profile requests from dependency events like AIPerf

  #564·dreamtalen opened Oct 22, 2026Contributor·

  ·
  9

• fix: refresh bundled zai-org GLM-5.3 configs to the published checkpoints

  #563·Arsene12358 opened Oct 22, 2026Contributor·

  ·
  2

• feat: accept native CUDA timings for autoregressive execution

  #561·milesial opened Oct 22, 2026Contributor·

  ·
  1

• feat: support session-pinned aggregated and disaggregated replay

  #560·milesial opened Oct 22, 2026Contributor·

  ·
  6

• feat: constrain recommendations by request latency p99

  #556·liangqianxing opened Oct 22, 2026·

  ·
  2

• fix: include context parallelism in the agg worker setup table

  #549·v0ropaev opened Oct 21, 2026·

  ·
  3

• fix: check saved CP rows against the generated config version

  #548·v0ropaev opened Oct 21, 2026·

  ·
  2

• ci: check bundled model config provenance

  #547·saturley-hall opened Oct 21, 2026Member·

  ·
  6

• fix: reject unsupported Task serving modes at construction

  #546·Arsene12358 opened Oct 21, 2026Contributor·

  ·
  8

• fix: reject unsupported pipeline parallelism in engine replay

  #545·milesial opened Oct 21, 2026Contributor·

  ·
  1

• fix: accept native dense KV skip evidence in FPM collection

  #529·simone-chen opened Oct 20, 2026Contributor·

  ·
  2

• fix: preserve load-match ranking across backends

  #524·git-jxj opened Oct 20, 2026·

  ·
  21

• feat: support vLLM G2 host offload with MTP

  #519·Harrilee opened Oct 19, 2026Contributor·

  ·
  7