All pull requests

Pull requests

• Add Llama family for the MLX backend

  #430·mikolaj92 opened Oct 5, 2026Contributor·

  ·
  1

• CUDA spill tier: GLM-5.3-Flash kept prompt states on local disk (--spill-gib)

  #427·wojo opened Oct 5, 2026·8 tasks done·

  ·

• feat(cuda): wire the spill tier into glm5_next serving (#373)

  #423·tfolkman opened Oct 5, 2026·

  ·
  1

• perf(cuda): drop the per-capture gc.collect from graph capture (#336)

  #422·tfolkman opened Oct 5, 2026·

  ·
  1

• GLM CUDA engine: decide which kept snapshots survive before cloning, so _take_over stops cloning snapshots it drops

  #421·m-naoki-m opened Oct 5, 2026·

  ·
  1

• i18n: Simplified Chinese for the command line and the control room

  #419·LXD-8 opened Oct 5, 2026·8 tasks done·

  ·

• feat(qwen4_exp): choose the MTP draft vocabulary, and ship a CJK-extended list

  #416·greatyingzi opened Oct 5, 2026·8 tasks done·

  ·
  1

• fix(tools): send a call to a tool the request did not offer under its own name (#256)

  #415·Alexbob0 opened Oct 5, 2026·

  ·
  1

• perf(flash next cuda): batch SSD n-gram gathers across streams

  #413·jregan-beasley-smc opened Oct 5, 2026·

  ·

• perf(kernels): the row decoder's window past 16 rows shares one threadgroup's weight reads

  #411·akol1 opened Oct 5, 2026Contributor·

  ·

• feat(glm cuda): EXL3 routed experts of mixed widths, a width per tensor from its header (#226)

  #410·MiaAI-Lab opened Oct 5, 2026Contributor·

  ·

• feat(deepseek-v41): add MLX family

  #409·akol1 opened Oct 5, 2026Contributor·

  ·

• cuda: draw top_k-off rows on the device (keyed, exact), with and without a top_p cut

  #408·sfxnz opened Oct 5, 2026·

  ·

• feat(metrics): /health's figures on /metrics: running tokens, cached prompt tokens, rounds, prefill time

  #407·MiaAI-Lab opened Oct 5, 2026Contributor·

  ·

• h3: re-voice a few-step clip with the base weights; 2x decoder; audio levels (draft, stacked on #384)

  #405·drowzeys opened Oct 5, 2026·

  ·

• tests(glm cuda): GLM-5.3-Flash's CUDA forward on the CPU, two ranks over gloo, against a float64 reference

  #404·MiaAI-Lab opened Oct 5, 2026Contributor·

  ·

• Flash Next CUDA: keep a state at the last 2048-row boundary before a system block's end, so a block with a new date line resumes there

  #403·grearjake-star opened Oct 4, 2026Contributor·

  ·

• feat: --dashboard, a live stats page

  #400·heitke opened Oct 4, 2026·

  ·

• feat(glm cuda): prompt_logprobs on completions

  #399·Bizuayeu opened Oct 4, 2026·

  ·

• perf(glm cuda): KDA prompt step kernel: eight warps a block, four warp sums in one (same bits)

  #398·Bizuayeu opened Oct 4, 2026·

  ·

• perf(glm cuda): a prompt chunk's rank exchanges in row pieces on a second stream (same bits)

  #397·Bizuayeu opened Oct 4, 2026·

  ·

• perf(glm cuda): DSA's prompt absorb/expand and pool scores on more programs (same bits)

  #396·Bizuayeu opened Oct 4, 2026·

  ·

• qwen_image: Qwen-Image-2.1 text to image on MLX (draft proposal, non-lane family, stacked on #384)

  #393·drowzeys opened Oct 4, 2026·

  ·

• v1.1: admission-time learning for the measured ladder (follow-up to #382) — rungs tighten from real solo fills

  #392·EugeneClaw opened Oct 4, 2026Contributor·

  ·
  1

• deepseek_v41: DeepSeek-V4.1-Flash on two DGX Sparks (EXL3 routed experts, DSpark drafts)

  #391·sfxnz opened Oct 4, 2026·

  ·
  2