All pull requests

Pull requests

• glm cuda: #2041's review: block 32's image only for csf0_v1, rows past 0x7E refused by name, a planted control, the form documented

  #2091·Bizuayeu opened Nov 4, 2026Contributor·8 tasks done·

  ·

• build: cuda.zig back to 600 lines (fixes red main)

  #2090·ashhart opened Nov 4, 2026Owner·

  ·

• cluster: serve --cluster brings a GB10 and a Mac up as one cluster over ConnectX and agrees on a disaggregated model's plan, before any tokens run (#1594, step 2a)

  #2088·ThinkOffApp opened Nov 4, 2026Contributor·9 tasks done·

  ·

• perf(hip): a skewed-row matrix tile, three blocks a WGP on gfx1151 for prompts up to 8192 codes deep

  #2087·AbrahamCete opened Nov 4, 2026Contributor·

  ·

• Qwen3.5: one layer-pattern reader for every backend (#1863, part 2)

  #2086·ashhart opened Nov 4, 2026Owner·

  ·
  1

• Metal: serve the dense Qwen3.8-27B straight from a GGUF file (Q4_K_M, Q8_0) (#1920)

  #2078·ashhart opened Nov 4, 2026Owner·

  ·

• Unlimited-OCR-2 on CUDA: the family's steps on NVIDIA GPUs, the router's long-context speedup (stacked on #2074)

  #2077·BobClawblaw opened Nov 4, 2026Contributor·8 tasks done·

  ·
  3

• Unlimited-OCR-2 on native Metal: the publisher's BF16 release, page images to structured text

  #2074·BobClawblaw opened Nov 3, 2026Contributor·

  ·
  1

• fix(cluster): refuse unplanned models in checks and serve

  #2069·swbiggart opened Nov 3, 2026Contributor·4 tasks done·

  ·
  1

• Living Weights on Gemma 4, step 2: the native GeGLU learner on CUDA and Metal (experimental, --slide refused)

  #2062·ashhart opened Nov 3, 2026Owner·

  ·

• fix(profile): report Nemotron CUDA paging overrides

  #2061·garvrao80 opened Nov 3, 2026Contributor·4 tasks done·

  ·
  2

• fabric stage: the protocol layer, boot description, round plan, rows layout and keep-alive, with golden bytes for every message (#2005 step 1b, stacked on #2012)

  #2060·jschmied opened Nov 3, 2026Contributor·

  ·
  5

• rope: Gemma 4 DFlash's drafter and Qwen-VL tower tables from core's helper (#1852, batches 4-5, value change)

  #2059·ashhart opened Nov 3, 2026Owner·

  ·
  1

• rope: Gemma 4's text and tower frequency tables from core's helper (#1852, batches 2-3, value change)

  #2058·ashhart opened Nov 3, 2026Owner·

  ·
  1

• Living Weights on Qwen4 dense, step 2: the native CUDA and Metal learner (experimental, --slide refused)

  #2054·ashhart opened Nov 3, 2026Owner·

  ·

• Clef-Flash: joint-schema decisions on /v1/systemone (HIP)

  #2048·laudney opened Nov 3, 2026Contributor·

  ·
  1

  4

• lanes: readout requests for decision prompts

  #2047·laudney opened Nov 3, 2026Contributor·

  ·
  6

• lfm3-vl: chat-template thinking suffix, fitGrid geometry, and deployment recipe

  #2045·viawright opened Nov 3, 2026·

  ·
  4

• glm cuda: decode windows' EXL3 experts run on the 16-pair item plan

  #2044·BadAd84 opened Nov 3, 2026Contributor·

  ·
  8

• feat(hip): one host protocol for the HIP OCR families

  #2040·laudney opened Nov 3, 2026Contributor·

  ·
  3

• Rerank: score a request's documents in one engine decision

  #2039·laudney opened Nov 3, 2026Contributor·

  ·
  5

• Vulkan: a dispatch records a barrier only when it conflicts with an earlier one, 3% to 7% less run time on an M1 with the same logits (stacked on #2017)

  #2038·joshuaswarren opened Nov 3, 2026Contributor·

  ·
  10

• CUDA: #1928's load read-ahead on Nemotron, Bolt, MiniMax M3, Kolibri-1, Spark 2.5 and DeepSeek-V4.1, opt-in (#2003)

  #2033·ashhart opened Nov 3, 2026Owner·

  ·

• DeepSeek-V4.1 TP=2: every-rank receipts and opt-in per-layer hash files for repeat divergence (#1395)

  #2030·ashhart opened Nov 3, 2026Owner·

  ·

• Living Weights: the 27B CUDA learner's SiLU-gate gradient diagnostics (#1806)

  #2029·ashhart opened Nov 3, 2026Owner·

  ·
  2