All pull requests

Pull requests

• feat(profiling): --profile-decode in-source decode profiler

  #156·ddvnguyen opened Sep 20, 2026Owner·

  ·
  1

• E0: measurement foundation (pin-load, timing hooks, lookup counters, gate)

  #152·ddvnguyen opened Sep 19, 2026Owner·

  ·

• WIP: gather-MMVQ pilot (review surface, not complete)

  #134·ddvnguyen opened Sep 18, 2026Owner·

  ·
  2

• docs(moe): expert tiering by execution site, not prediction

  #131·ddvnguyen opened Sep 16, 2026Owner·

  ·

• MoE look-ahead PR-P1: producer (NOT LANDABLE, see body)

  #127·ddvnguyen opened Sep 13, 2026Owner·

  ·
  1

• cuda : MoE look-ahead expert prefetch (PR-A consumer) + LFRU eviction guard

  #126·ddvnguyen opened Sep 13, 2026Owner·

  ·

• docs : decode MoE look-ahead expert prefetch design

  #125·ddvnguyen opened Sep 13, 2026Owner·

  ·

• cuda : add opt-in MoE cache telemetry env and early-router diagnostics

  #123·ddvnguyen opened Sep 13, 2026Owner·

  ·

• example : add MoE route-trace for offline routing analysis

  #122·ddvnguyen opened Sep 13, 2026Owner·

  ·

• Reconcile qwen4exp NextN/MTP draft head: adopt gs qwen4exp-mtp lineage over hand-rolled cdd11021b (#763)

  #120·ddvnguyen opened Sep 12, 2026Owner·

  ·
  9

• arm: uniform q8_0 KV (main+draft) vs production mixed pin — measured loss on the rig

  #119·ddvnguyen opened Sep 11, 2026Owner·

  ·

• feat(cuda): arm117 — port adaptive-KV-streaming onto baseline + Gate G/R/B pass; rig cells pending (#117)

  #118·ddvnguyen opened Sep 11, 2026Owner·

  ·
  2

• arm: adaptive-KV-streaming correctness/perf on the 2x-RTX RPC-split topology (design phase)

  #117·hydra-z[bot] opened Sep 10, 2026Bot·

  ·

• Arm spec: context-shift hybrid-state correctness for Qwen3.8 (design phase)

  #116·ddvnguyen opened Sep 10, 2026Owner·

  ·
  1

• feat(cuda): PR103.0 GDN cache-cpy fusion — A/B toggle + test coverage (implements #112)

  #115·ddvnguyen opened Sep 10, 2026Owner·

  ·

• arm PR106.x: EXL3 DFlash2 kit reference series (PR106.0 external / PR106.1 native 3.5bpw single-GPU)

  #114·ddvnguyen opened Sep 9, 2026Owner·

  ·

• arm PR104.x: mixed-quant row sharding arm series (PR104.0 roofline / PR104.1 spike / PR104.2 production)

  #113·ddvnguyen opened Sep 9, 2026Owner·

  ·

• arm PR103.0: CUDA GDN cache-cpy fusion (arm spec + implementation PR spec)

  #112·ddvnguyen opened Sep 9, 2026Owner·

  ·

• baseline arm PR105.0: single-machine dual-GPU (no RPC), PR #110 config reference

  #111·ddvnguyen opened Sep 9, 2026Owner·

  ·
  1

• fix(#498): package all shared libs in sm60 OCI image (ldd verify, LD_LIBRARY_PATH)

  #107·ddvnguyen opened Sep 2, 2026Owner·

  ·
  1

• fix(713): EAGAIN retry helper + slot quarantine on STATE_PUT restore failure

  #106·ddvnguyen opened Aug 29, 2026Owner·

  ·
  1

• epic(610): extract Hydra server logic out of server-context.cpp (mechanical, A/B toggle)

  #105·ddvnguyen opened Aug 21, 2026Owner·

  ·

• fix(470): auto-recovery probe + force T3 rebuild on RPC peer reconnection

  #104·hydra-z[bot] opened Aug 21, 2026Bot·

  ·

• ci: hydra-engine PR validation (build+ctest, server pytest, C# parity)

  #81·ddvnguyen opened Aug 6, 2026Owner·

  ·