All pull requests

Pull requests

• maya (DeepSeek setup): ISTA-DASLab's GSQ-RCO Coder GGUF as a listed quant

  #964·ImCodingCat opened Nov 12, 2026·

  ·

• DeepSeek: the RAM store's slots a layer's size, so a UD quant's store holds as many experts as a copy (decode +20% over the copy, the same bits)

  #963·mrsirg97-rgb opened Nov 12, 2026Contributor·

  ·

• Kimi K3.1 Pro support (maya-k31): the whole model on three 24 GB GPUs plus an NVMe drive

  #962·captcrock opened Nov 11, 2026·

  ·

• GLM: keep every expert the tiers already paid for, so a spilled GLM-5.3 reads less from the SSD (decode +21%, the same bits)

  #960·mrsirg97-rgb opened Nov 11, 2026Contributor·

  ·

• DeepSeek: a short prompt's least-picked RAM experts on the CPU lane while the GPU streams the rest (MAYA_DS4_PREFILL_CPU, opt-in) - an agent's turn ~40% faster on PCIe 3.0

  #959·PrimusInterParis opened Nov 11, 2026·

  ·

• DeepSeek: DS4_FLOW_COPY_DMA also where the RAM copies are split across CPU sockets - short prompts 11-16% faster on two sockets, the same bits

  #958·PrimusInterParis opened Nov 11, 2026·

  ·

• DeepSeek: allow ordered CPU-lane worker affinity

  #953·handmade0octopus opened Nov 11, 2026Contributor·

  ·
  2

• Tesla P100 (sm_60): decode and prompt kernels that suit the card - 4x P100, Maya-S24: decode +13%, prompts +34-53%

  #952·KaneQuah opened Nov 11, 2026·

  ·

• Server: POST /v1/completions (raw prompt, optional fill-in-the-middle suffix) for editor autocomplete

  #950·talisp opened Nov 11, 2026·

  ·

• glm: add opt-in NUMA CPU-list setup with user-selected threads

  #949·oscarlius opened Nov 11, 2026Contributor·

  ·

• DeepSeek: the Engram row cache sized by the RAM the placed experts leave, at most 16 GiB of rows (was 256 MB at most)

  #947·olobarius opened Nov 10, 2026·

  ·

• HIP EXL3 decode: mv_kernel inlines block_gemv; the trellis shifts and codebook 2 multiply at full rate (same bits)

  #945·pepuscz opened Nov 10, 2026Contributor·

  ·

• Preserve prompt state on interrupted Metal requests

  #941·appleguy opened Nov 10, 2026Contributor·

  ·
  1

• docs: warn about uneven GGUF TP on PCIe

  #940·mercant1982 opened Nov 10, 2026·

  ·

• docs: add 4x RTX 4090 GLM-5.3-Flash Maya-L benchmark to speed table

  #938·oscarlius opened Nov 10, 2026Contributor·

  ·

• DeepSeek: retain prompt DMA pins when decode streaming turns off

  #937·NektoDron opened Nov 9, 2026·

  ·

• DeepSeek: discard failed GGUF expert-cache moves

  #935·NektoDron opened Nov 9, 2026·

  ·
  3

• DeepSeek: build a Maya drafter pack from an existing DSpark GGUF

  #933·mercant1982 opened Nov 9, 2026·

  ·

• DeepSeek: export a kept conversation and import it into another server (POST /slots/export, /slots/import; step 1 of #655)

  #932·mrsirg97-rgb opened Nov 9, 2026Contributor·

  ·

• ds4: accelerate MXFP4 CPU experts on pre-AVX2 x86 CPUs

  #931·UnagiDojyou opened Nov 9, 2026·

  ·
  2

• DeepSeek EXL3: tools/ds4/reorder.py rewrites type-ordered shards in expert order (Diffbot's 3 bpw, #806)

  #929·gopinath87607 opened Nov 9, 2026Contributor·

  ·
  5

• ds4 CPU lane: one flowing batch per call, the own multi-token IQ1_S rows on multi-GPU boxes, and the bit-exact rows tests

  #928·saatvik-viniak opened Nov 9, 2026Contributor·

  ·
  3

• WebUI Chat: Safari on macOS submits the message when pressing Enter to confirm Japanese IME conversion

  #927·takahiro-blab opened Nov 9, 2026·

  ·
  1

• Setup: recognise AMD cards in two card-size checks

  #923·Mqrius opened Nov 8, 2026·

  ·
  2

• DeepSeek: support image UUID references

  #922·NektoDron opened Nov 8, 2026·

  ·
  6