All pull requests

Pull requests

• nemotron omni: video input through the RADIO tower

  #940·MiaAI-Lab opened Oct 16, 2026Contributor·8 of 9 tasks·

  ·

• drama3: Fish Audio Drama 3 text to speech on one CUDA GPU (new family, plain decoding; proposal)

  #939·drowzeys opened Oct 16, 2026·

  ·

• nemotron: serve NVIDIA Nemotron 3.5 Bolt (nemotron_h_puzzle) on CUDA

  #938·MiaAI-Lab opened Oct 16, 2026Contributor·6 of 9 tasks·

  ·

• nemotron metal: a new build keeps the chip's last window-cost table while a probe holds

  #936·jishnuvenugopal opened Oct 16, 2026Contributor·9 tasks done·

  ·
  3

• KV pages: one core owner with reservation leases, Nemotron CUDA pilot (#781)

  #935·jschmied opened Oct 16, 2026Contributor·

  ·
  1

• cuda tests: reject non-finite errors before reducing FP64 comparisons

  #933·cwschroeder opened Oct 16, 2026Contributor·

  ·

• perf(qwen3_5 cuda): deliver the first token before drafting

  #932·jregan-beasley-smc opened Oct 16, 2026·

  ·

• app: hosted models carry the context window their endpoint reports

  #930·Riven-Spell opened Oct 16, 2026·

  ·
  2

• Disable Nagle on accepted HTTP streaming connections

  #929·akol1 opened Oct 16, 2026Contributor·

  ·

• fix(lanes): a pipelined stream adopted between rounds starts in verify

  #928·chaog992 opened Oct 16, 2026Contributor·

  ·

• drama: the backbone chain launches with programmatic dependent launch

  #927·plotarmordev opened Oct 16, 2026Contributor·4 of 8 tasks·

  ·

• drama: the tower's one-row projections on drama_tower_gemv, in cuBLASLt's summation order

  #926·plotarmordev opened Oct 16, 2026Contributor·4 of 8 tasks·

  ·

• drama: load the checkpoint with direct I/O in file order and build the codec on its own thread

  #925·plotarmordev opened Oct 16, 2026Contributor·4 of 8 tasks·

  ·

• drama: faster tower decode attention and sampling, same bits

  #924·plotarmordev opened Oct 16, 2026Contributor·4 of 8 tasks·

  ·

• drama: faster one-row backbone projections and glue, same bits

  #923·plotarmordev opened Oct 16, 2026Contributor·4 of 8 tasks·

  ·

• speech: a Drama 3 reply that reaches max_frames returns its clipped audio

  #922·plotarmordev opened Oct 16, 2026Contributor·4 of 8 tasks·

  ·

• speech: streamed Drama 3 audio on /v1/audio/speech

  #921·plotarmordev opened Oct 16, 2026Contributor·4 of 8 tasks·

  ·

• drama: Drama 3 (fishaudio/drama-3) text-to-speech on CUDA

  #920·plotarmordev opened Oct 16, 2026Contributor·4 of 8 tasks·

  ·

• Reject truncated tensors in the Flash Next pack comparator

  #919·cwschroeder opened Oct 16, 2026Contributor·

  ·
  2

• cuda: Module.holds, whether a loaded image defines a device global

  #917·plotarmordev opened Oct 16, 2026Contributor·6 of 8 tasks·

  ·

• core: a direct read the file system refuses goes through the page cache

  #916·plotarmordev opened Oct 16, 2026Contributor·6 of 8 tasks·

  ·

• cuda: a process-wide cap on DeviceBuffer and pinned HostBuffer bytes

  #915·plotarmordev opened Oct 16, 2026Contributor·6 of 8 tasks·

  ·

• tool_parse: unclosed namespaced tool-call openers no longer rescan the reply

  #913·jishnuvenugopal opened Oct 16, 2026Contributor·8 tasks done·

  ·

• Report unavailable engine health and preserve admission HTTP 503

  #912·akol1 opened Oct 16, 2026Contributor·

  ·
  1

• Living Weights on Qwen3.8-27B (CUDA, GB10): the pocket in every pass, the trainer, --slide, and pocket neurons that no longer drive one another

  #911·ZackSample opened Oct 16, 2026Contributor·

  ·
  1