All pull requests

Pull requests

• feat(spyre): tensor parallelism with card-to-card collectives at two, four and eight ranks

  #1019·starpit opened Nov 1, 2026Contributor·11 of 18 tasks·

  ·
  2

• feat(metal): draft Qwen3.6-35B-A3B tokens with Splash's DFlash2 drafter, compiled ahead of time

  #1015·starpit opened Oct 31, 2026Contributor·6 of 29 tasks·

  ·
  23

• feat(spyre): torch-spyre's paged attention, ported to Triton and run on the emulator

  needs-fix
  #999·starpit opened Oct 29, 2026Contributor·

  ·
  219

• test(spyre): pin the sampler's all-NaN sentinel and the readback's refusal of it

  lgtm
  #985·starpit opened Oct 29, 2026Contributor·

  ·
  1

  1

• feat(cuda): rebuild the CUDA backend on ThunderKittens for H100 (Qwen3.5 0.8B/9B, Qwen3.6-35B-A3B, Gemma 4 26B-A4B)

  target/cuda
  #978·diegocastanibm opened Oct 28, 2026Collaborator·

  ·

• feat(bench): coding workloads, Polyglot Python scored by its tests in a sandbox and SWE-bench Lite prompts

  enhancement
  #954·aavarghese opened Oct 27, 2026Collaborator·

  ·

• fix(metal): store each TurboQuant vector's norm over its codes' decoded length

  lgtm
  #940·starpit opened Oct 27, 2026Contributor·

  ·

• feat(site): the metal overview and Over time follow the dataset workloads

  enhancementlgtm
  #929·aavarghese opened Oct 26, 2026Collaborator·

  ·
  1

• fix(spyre): paged Triton attention answers the oracle on card

  target/spyre
  #928·starpit opened Oct 26, 2026Contributor·

  ·

• feat(bench): the metal matrix times real dataset tasks by default, and the page shows them

  enhancementlgtm
  #920·aavarghese opened Oct 26, 2026Collaborator·

  ·
  4

• ci(kani): run one Kani proof per crate on Kani 0.68, whose nightly meets ktir-core's rust-version

  needs-fix
  #915·starpit opened Oct 25, 2026Contributor·

  ·
  1

  1

• test(tensors): make cfg_neutrality read multi-line cfg attributes, so a rustfmt wrap cannot hide a target gate

  lgtm
  #911·starpit opened Oct 25, 2026Contributor·

  ·
  1

• feat(bench): add cc_trace.py live, the Lane B runner

  #835·diegocastanibm opened Oct 22, 2026Collaborator·

  ·
  3

• chore(cuda): remove the current CUDA implementation

  #721·diegocastanibm opened Oct 20, 2026Collaborator·

  ·
  1

• feat(cli): send Claude Code's auto-mode safety checks to local scratchy

  #713·diegocastanibm opened Oct 19, 2026Collaborator·

  ·
  17

• feat(models): gpt-oss-20b speculative decoding with AWS's P-EAGLE drafter

  #602·starpit opened Oct 17, 2026Contributor·6 of 13 tasks·

  ·
  1

• docs(bench): record what to expect from the KV-cache ablation (T0.8)

  documentation
  #275·diegocastanibm opened Oct 7, 2026Collaborator·

  ·
  2

• docs(bench): pin R2 to a byte-verified pair after the -mlx tag audit (T0.6)

  documentation
  #257·diegocastanibm opened Oct 6, 2026Collaborator·6 tasks done·

  ·
  5

• feat(site): add the Launch Claude Leaderboard page, schema and validator

  #176·diegocastanibm opened Oct 1, 2026Collaborator·8 of 12 tasks·

  ·

• feat(chat-template): generate the renderer from a format DSL we own

  #56·diegocastanibm opened Sep 17, 2026Collaborator·

  ·
  2