All pull requests

Pull requests

• chore(bench): add cuBLASLt version and gemv A/B harness for GB10

  priority:lowstatus:in-progresstype:chore
  #2775·inureyes opened Oct 30, 2026Member·

  ·

• docs(benchmarks): add full models/ sweep for ROCm gfx1151

  area:benchmarkplatform:linuxpriority:lowstatus:donetype:docs
  #2774·inureyes opened Oct 30, 2026Member·

  ·

• fix(vision): use the transformers mask for Gemma 3 image prefill

  area:modelsmodeltype:vlmpriority:mediumstatus:reviewtype:bug
  #2773·inureyes opened Oct 30, 2026Member·5 tasks done·

  ·
  1

  2

• fix(router): re-probe down prefill peers before answering 503

  area:routerpriority:lowstatus:donetype:bug
  #2772·inureyes opened Oct 30, 2026Member·4 tasks done·

  ·
  1

  2

• fix(media): cap DNS per domain, redact hop schemes, check NAT64 registry

  priority:mediumstatus:donetype:security
  #2771·inureyes opened Oct 30, 2026Member·

  ·
  1

  2

• feat(speculative): run an ordinary full model as a drafter on the engine

  area:corepriority:highstatus:reviewtype:enhancement
  #2768·inureyes opened Oct 30, 2026Member·6 of 7 tasks·

  ·
  1

• update(cuda): fold batched decode rows in QuantizedMatmul

  area:coreplatform:windowspriority:highstatus:in-progresstype:performance
  #2767·inureyes opened Oct 30, 2026Member·

  ·
  1

• refactor(speculative): remove the classic draft-model path

  area:corepriority:highstatus:reviewtype:refactor
  #2756·inureyes opened Oct 30, 2026Member·

  ·
  1

  2

• [WIP] rocm cdna warp size qwen35 mrope batch

  #2731·daemyung-lablup opened Oct 29, 2026Contributor·

  ·

• docs(reports): generate an index for TECHNICAL_REPORTS

  priority:lowstatus:reviewtype:docs
  #2689·inureyes opened Oct 27, 2026Member·

  ·
  1