All pull requests

Pull requests

• Living Weights: facts learned and checked in conversation, not only asked cold

  #891·ZackSample opened Oct 15, 2026Contributor·8 tasks done·

  ·

• Support optional bearer authentication in benchmark clients

  #890·alenperic opened Oct 15, 2026·7 of 9 tasks·

  ·

• Fix incomplete streams counted as successful in shared-prefix load tests

  #889·alenperic opened Oct 15, 2026·7 of 9 tasks·

  ·

• Qwen3.5/3.8 on CUDA: kept prompt states in host memory on #836's engine (#726's reuse; on #885)

  #888·BobClawblaw opened Oct 15, 2026Contributor·

  ·
  1

• Qwen3.8-27B decode: K splits measured over every layer's weights, and an undrafted one-row round kept by its verify

  #887·BobClawblaw opened Oct 15, 2026Contributor·

  ·
  1

• qwen3_5 cuda: --kv-cache fp8, e4m3 attention caches with a scale a row

  #886·BobClawblaw opened Oct 15, 2026Contributor·8 tasks done·

  ·
  1

• Qwen3.5/3.8 on CUDA: image input on #836's engine (2B, 27B; on #836 and #853)

  #885·BobClawblaw opened Oct 15, 2026Contributor·

  ·
  2

• tool_parse: a GLM call's keys without values read in linear time

  #883·jishnuvenugopal opened Oct 15, 2026·8 tasks done·

  ·
  1

• server: a reply of many tool calls reads in linear time

  #882·jishnuvenugopal opened Oct 15, 2026·8 tasks done·

  ·
  1

• tool_parse: a Gemma call's strings go back in one pass

  #881·jishnuvenugopal opened Oct 15, 2026·8 tasks done·

  ·
  1

• Gemma 4 31B served natively on Metal: the dense family, DFlash drafts, tensor-unit prompts past mlx-lm, and core affine_mm without its epilogue

  #879·BobClawblaw opened Oct 15, 2026Contributor·

  ·
  1

• Living Weights as a core module: the learner, pocket and fits in zig/src/living, a family described as data (Nemotron only)

  #878·ZackSample opened Oct 15, 2026Contributor·8 tasks done·

  ·
  1

• Living Weights: every earlier fact asked again after a lesson, and a take-out that puts the pocket back bit for bit

  #877·ZackSample opened Oct 15, 2026Contributor·8 tasks done·

  ·
  1

• Living Weights: a quiet refit holds an earlier fact's neuron on its own questions as read now

  #876·ZackSample opened Oct 15, 2026Contributor·8 tasks done·

  ·
  1

• nemotron metal: a failed load releases the kernels it compiled

  #875·jishnuvenugopal opened Oct 15, 2026·8 tasks done·

  ·

• ling: Ling-3.1-flash on CUDA at TP=1 on #804's kernels (draft, stacked on #804)

  #874·jschmied opened Oct 15, 2026Contributor·6 of 8 tasks·

  ·
  4

• perf(lfm2): prompt calls on the M5 tensor units

  #873·BlivionIaG opened Oct 15, 2026Contributor·8 tasks done·

  ·

• Living Weights: delete one learned fact, recheck the rest, put it back on failure

  #872·ThinkOffApp opened Oct 15, 2026Contributor·6 of 8 tasks·

  ·

• perf(core): affine_mm 8-byte tile loads on M5 and mlx-lm's plain SwiGLU

  #871·BlivionIaG opened Oct 15, 2026Contributor·8 tasks done·

  ·

• Flash Next on CUDA: the qwen4_exp family on the native CUDA host, every stream through the lane rounds

  #869·ZackSample opened Oct 15, 2026Contributor·7 of 8 tasks·

  ·
  3

• perf(lfm2): windows of 16 rows or more read each picked expert once

  #868·BlivionIaG opened Oct 15, 2026Contributor·8 tasks done·

  ·
  3

• perf(core): tile_qmm window twins, one weight read for the rows of a window

  #867·BlivionIaG opened Oct 15, 2026Contributor·8 tasks done·

  ·
  6

• docs: the native /health fields, and its CUDA counters marked as the 0.6.6 server's (#753)

  #866·jishnuvenugopal opened Oct 15, 2026·8 tasks done·

  ·
  1

• CUDA: retain pinned peaks and scoped process memory receipts

  #865·jayleaton opened Oct 15, 2026Contributor·4 of 8 tasks·

  ·
  1

• GLM-5.3-Flash: end the last expert's sorted pairs at the pair count

  #864·harrisonfriia opened Oct 15, 2026Contributor·8 tasks done·

  ·
  1