All pull requests

Pull requests

• GLM-5.3-Flash faster oQ kernels in Zig engine

  #786·grantoverton opened Oct 13, 2026·8 tasks done·

  ·

• Qwen3.8-27B faster oQ kernels in Zig engine

  #784·grantoverton opened Oct 13, 2026·8 tasks done·

  ·

• perf(nemotron cuda): the lone driver reclaims the lane rounds' last stream

  #780·plotarmordev opened Oct 13, 2026Contributor·7 of 8 tasks·

  ·

• perf(nemotron cuda): the lone driver takes structured output

  #779·plotarmordev opened Oct 13, 2026Contributor·7 of 8 tasks·

  ·

• perf(nemotron cuda): the lone driver takes thinking budgets and loop guards

  #778·plotarmordev opened Oct 13, 2026Contributor·7 of 8 tasks·

  ·

• perf(nemotron cuda): the lone driver takes requests with logprobs

  #777·plotarmordev opened Oct 13, 2026Contributor·7 of 8 tasks·

  ·

• perf(nemotron cuda): prompt_logprobs reads 256 rows a head pass

  #776·plotarmordev opened Oct 13, 2026Contributor·8 tasks done·

  ·

• nemotron metal: overlap and preload taught-pocket decode work

  #775·ThinkOffApp opened Oct 13, 2026Contributor·6 of 8 tasks·

  ·

• Qwen3.8-Flash-Next support for oQ in Zig engine

  #774·grantoverton opened Oct 13, 2026·8 tasks done·

  ·
  1

• GLM-5.3-Flash support for oQ in Zig engine

  #773·grantoverton opened Oct 13, 2026·8 tasks done·

  ·
  1

• Nemotron CUDA: decode glue keeps its loads in flight, same values and order

  #771·plotarmordev opened Oct 13, 2026Contributor·8 tasks done·

  ·

• Nemotron CUDA: a shared round's per-stream conv, scan and attention run as one launch each

  #770·plotarmordev opened Oct 13, 2026Contributor·8 tasks done·

  ·

• lanes: a shared round prices drafts by their head steps, a batched head by the deepest stream's

  #769·plotarmordev opened Oct 13, 2026Contributor·8 tasks done·

  ·

• Nemotron CUDA: a shared round's MTP heads draft every stream in one batch

  #768·plotarmordev opened Oct 13, 2026Contributor·7 of 8 tasks·

  ·

• server: the tokenizer keeps long text pieces' ids, so later turns encode only their new text

  #767·plotarmordev opened Oct 13, 2026Contributor·8 tasks done·

  ·

• core: a checkpoint's pages out of the page cache once a loader is done with them

  #766·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·

• glm cuda: agree settings across every rank and give each follower its own doorbell

  #765·Kurcide opened Oct 13, 2026·

  ·

• cuda: the staged NVFP4 prompt entries leave an expert's items alone (skip), as the decode form does

  #764·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·

• cuda: Nvfp4Experts.packGateUp packs gate and up in place, no scratch

  #763·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·

• native: a CUDA family's prompt cache, weight share and open errors come from its own declarations

  #762·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·

• server: GLM tool calls the model wrote with a key or an end missing still parse

  #761·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·

• server: the soft open-file limit is raised to the hard one when the server listens

  #760·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·

• server: Responses accepts OpenAI's include values and ignores them

  #759·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·

• core: read ModelOpt NVFP4 checkpoints as stored

  #758·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·
  4

• cuda: the NVFP4 lane matmul and prompt GEMM write fp32 too (qlinear decodeF32 and promptF32)

  #757·Bizuayeu opened Oct 13, 2026Contributor·8 tasks done·

  ·