All pull requests

Pull requests

• fix(web): restore the reasoning stream the redesign dropped

  #1697·kevin9327 opened Sep 22, 2026·

  ·

• qwen36: keep the prefill echo out of the segment build

  #1695·namespaceMarcello opened Sep 22, 2026·

  ·

• fix(olmoe): wait for an expert already being read instead of reading it twice

  #1693·namespaceMarcello opened Sep 22, 2026·

  ·

• release 1.12.1: dev to main

  #1663·JustVugg opened Sep 21, 2026·

  ·

• qwen38: explain parallel expert-read fallbacks

  #1646·Stamina9 opened Sep 21, 2026·

  ·
  2

• cuda: decode fmt=9 (bf16) in the dense matmul

  #1620·DebugSultan opened Sep 19, 2026·

  ·
  2

• fix(olmoe): use the shared RAM probe, add a plausibility floor, warn …

  #1611·Cometbuster4969 opened Sep 18, 2026·

  ·
  1

  3

• Add optional desktop launcher for Windows and Linux

  #1607·ipechman opened Sep 18, 2026·

  ·

• dense Qwen3 family: engine, int4-gs64 converter, tiny-fixture oracle, registry entry

  #1578·Xore opened Sep 16, 2026·

  ·

• perf(olmoe): O(1) LRU victim selection via intrusive recency lists (#1050)

  #1571·Petsku01 opened Sep 16, 2026·

  ·
  2

• qt expert tier + expert-cache RAM ceilings (glm53 & qwen36), A3 soak-verified

  #1564·Xore opened Sep 16, 2026·

  ·

• qwen38: the int8 trunk carries one scale per 64 weights (grouped fmt 1 on the GPU), perplexity-neutral

  #1552·kreuzzelg opened Sep 15, 2026·

  ·
  1

• fix(windows): load CUDA backend before device discovery

  #1542·IcarusAegis opened Sep 15, 2026·

  ·
  5

• feat(glm53): add device-resident HIP backend for MI350P

  #1448·azharlabs opened Sep 12, 2026·

  ·

• perf: COALESCE=1 multi-expert coalesced pread for contiguous shard runs

  #1398·KyleSanderson opened Sep 8, 2026·

  ·

• perf: IDOT_TEAM=1 fuses three OpenMP regions into one team per expert…

  #1397·KyleSanderson opened Sep 8, 2026·

  ·
  2

• Opt-in exact verify mode for speculative batches (COLI_EXACT_VERIFY=1)

  #1395·rybruscoe opened Sep 7, 2026·

  ·
  2

• feat(engine): replace the ablation scoring mode with a checked, digest-bound one

  #1356·monotophic opened Sep 5, 2026·

  ·
  4

• OpenAI API Instrumentation (e.g. seed, logprobs, echo, an array prompt)

  #1353·monotophic opened Sep 5, 2026·

  ·
  4

• feat(qwen36): load MLX-affine dense weights and norms from a qpack container

  #1343·Avicennasis opened Sep 4, 2026·

  ·
  3

• qwen36: Vulkan expert tier, and staged device-local uploads for cards without Resizable BAR

  #1338·crichalchemist opened Sep 4, 2026·

  ·
  8

• feat(qwen36): routed experts from a qpack container through bounded Metal slots

  #1323·Avicennasis opened Sep 2, 2026·

  ·
  5

• fix(planner): honor cgroup memory limits

  #1316·Avicennasis opened Sep 1, 2026·

  ·
  4

• Add resumable, hash-verified qpack installers for Hugging Face and static mirrors

  #1315·Avicennasis opened Sep 1, 2026·

  ·
  5

• perf(sse41): vectorize grouped-int4 GEMV

  #1286·cameron opened Aug 30, 2026·

  ·
  2