All pull requests

Pull requests

• [Perf][PersonaPlex] Enable backbone decode prefill graph

  #2607·WilsonZheng0327 opened Oct 7, 2026Collaborator·4 of 5 tasks·

  ·

• [Docs] Add Qwen3-TTS variants and AuK XPU launch recipes

  #2606·yao-matrix opened Oct 7, 2026Contributor·3 tasks done·

  ·
  6

• [Pipeline] Stop stage workers when the server process dies

  run-ci
  #2605·rwang5203 opened Oct 7, 2026Collaborator·

  ·

• [Fix] Keep matched stop strings out of chat and /generate text

  run-ci
  #2604·rwang5203 opened Oct 7, 2026Collaborator·

  ·

• [Fix] Keep the dLLM scheduler serving when a request fails to build

  run-ci
  #2603·rwang5203 opened Oct 7, 2026Collaborator·

  ·

• [Fix] Reject Qwen3-Omni prompt text that cannot fit before tokenizing it

  run-ci
  #2602·rwang5203 opened Oct 7, 2026Collaborator·

  ·

• [Perf][Nemotron-VoiceChat] Admit Talker requests to a busy batch only once their text arrives

  #2601·XinhaoTheo opened Oct 7, 2026Contributor·3 of 5 tasks·

  ·

• [Fix] Apply the media policy to request audio of ASR, PersonaPlex and Nemotron VoiceChat

  run-ci
  #2600·rwang5203 opened Oct 7, 2026Collaborator·

  ·

• [Qwen3-Omni Perf] compile the thinker's MoE kernel variants and the talker's repetition penalty at load

  run-ci
  #2599·Ratish1 opened Oct 7, 2026Collaborator·

  ·
  1

• [Qwen3-Omni Perf] overlap the fused predictor launches and fold each codebook's argmax, gather and add

  run-cirun-qwen3-tts
  #2598·Ratish1 opened Oct 7, 2026Collaborator·

  ·
  1

• [Qwen3-Omni Perf] ship the talker only the encoder metadata its MRoPE reads

  run-ci
  #2597·Ratish1 opened Oct 7, 2026Collaborator·

  ·
  3

• [Perf][PersonaPlex] Continuously batch streaming Mimi decode

  #2596·nightdecode opened Oct 7, 2026·3 of 5 tasks·

  ·

• [Perf][MiniCPM-o] Avoid SHM for same-process image encoder inputs

  #2595·shaoxiawjc opened Oct 7, 2026·

  ·
  1

• [Perf][PersonaPlex] Fuse Depformer pointwise ops with CuTe DSL

  #2594·LiuYuHui opened Oct 7, 2026·5 tasks·

  ·

• feat(voxt): native Cohere Transcribe 03-2026 runtime

  #2593·khazic opened Oct 7, 2026Collaborator·2 of 5 tasks·

  ·

• feat(voxt): native Whisper small fp16 support

  #2592·khazic opened Oct 7, 2026Collaborator·1 of 4 tasks·

  ·

• feat(voxt): native Whisper large-v3 fp16 support

  #2591·khazic opened Oct 7, 2026Collaborator·1 of 4 tasks·

  ·

• feat(voxt): native Whisper large-v3-turbo runtime

  #2590·khazic opened Oct 7, 2026Collaborator·1 of 4 tasks·

  ·
  1

• [Perf][MiniCPM-o][Video] Parallelize video checkpoint preprocessing

  #2589·charliechenye opened Oct 7, 2026Collaborator·4 of 5 tasks·

  ·
  1

• [Voxt] Run Sortformer v2.1 speaker diarization on the native runtime

  run-ci
  #2588·JiaxinD opened Oct 7, 2026Collaborator·4 tasks done·

  ·

• [Voxt] Run Silero VAD v6 on the native runtime

  run-ci
  #2587·JiaxinD opened Oct 7, 2026Collaborator·4 tasks done·

  ·

• [Voxt] Run Qwen3-ASR 1.7B 6bit and 8bit on the native runtime

  run-ci
  #2586·JiaxinD opened Oct 7, 2026Collaborator·4 tasks done·

  ·

• [Fix][Nemotron-VoiceChat] Stop the Talker input rows leaking GPU memory

  #2585·XinhaoTheo opened Oct 7, 2026Contributor·3 of 4 tasks·

  ·

• [Voxt] Native C++ Qwen3-ASR runtime on MLX (no Python runtime)

  run-ci
  #2577·JiaxinD opened Oct 6, 2026Collaborator·4 tasks done·

  ·

• [Perf] Support seeded sampling under thinker async decode

  #2573·YifanLi3 opened Oct 6, 2026·3 of 5 tasks·

  ·