All pull requests

Pull requests

• docker image cuda build fix + cuda12 extension

  #1848·rafal-prasal opened Oct 10, 2026Contributor·

  ·

• serve: an opt-in VRAM headroom for --expert-cache auto (#1549)

  #1841·demetree opened Oct 10, 2026Contributor·

  ·

• serve: a reply that ends inside its thinking with no answer is continued (#1814)

  #1840·47Hunter47 opened Oct 10, 2026·

  ·
  1

  1

• fix(serve): preserve literal closing thinking tags

  #1839·CC-David-CC opened Oct 10, 2026Contributor·

  ·

• Resident RAM mode: the RAM copy follows the conversation (STRATA_RAM_ADAPT, opt-in)

  #1838·SanghunYun opened Oct 10, 2026·

  ·
  1

• serve: a Chinese / English language switch for the web app (adapted to v0.1.41)

  #1836·hogtinbao opened Oct 10, 2026·

  ·

• Community benchmark: 2x RTX PRO 4500 Blackwell, Swift IQ3_XXS, engine 0.1.41 against 0.1.40.1 (single stream to 248k, --batch 8 sweep)

  #1834·qni-live opened Oct 10, 2026Contributor·

  ·

• Add optional isolated TeX Live rendering to web chat

  #1833·CC-David-CC opened Oct 10, 2026Contributor·

  ·

• Add locally bundled MathJax rendering to web chat

  #1832·CC-David-CC opened Oct 10, 2026Contributor·

  ·

• Add locally bundled KaTeX rendering to web chat

  #1830·CC-David-CC opened Oct 10, 2026Contributor·

  ·

• Add shared LaTeX/web math core without a rendering engine

  #1829·CC-David-CC opened Oct 10, 2026Contributor·

  ·

• sycl: register iq_parity and ple_parity like the top-level build

  #1828·Momoyeyu opened Oct 10, 2026·

  ·
  1

• qsa_select: prompt-path block scores with a thread per query head (below sm_80, same bits)

  #1825·xrip opened Oct 10, 2026·

  ·

• cpu: Q2_0 rows on AVX-only CPUs (q2_avx1.cpp) instead of ggml-cpu's scalar dot

  #1824·xrip opened Oct 10, 2026·

  ·

• tokenizer: a long message's ids remembered whole, on top of the piece cache

  #1823·xrip opened Oct 10, 2026·

  ·

• prefill: gate/up MMQ reads the experts in place (CUDA; patched copy of llama.cpp mmq.cuh)

  #1822·xrip opened Oct 10, 2026·

  ·

• qsa_prompt_attn: int8 K/V prefetched in registers and sharing one buffer (Turing / sm_75)

  #1821·xrip opened Oct 10, 2026·

  ·

• qwen35moe follow-ups: VRAM cap + quality mode, MTP for F16 draft tensors, chunked GDN, 256/8 route-resident

  #1820·atomicmilkshake opened Oct 10, 2026·

  ·

• fix(cpu): make IQ expert dispatch independent of draft grouping

  #1818·CC-David-CC opened Oct 10, 2026Contributor·

  ·

• sycl: use the source version for setup model checks

  #1817·nekomario28 opened Oct 10, 2026·3 tasks done·

  ·
  1

• Opt-in OTLP traces of every /v1 request (0.1.41), plus the docs pages 0.1.41 left uncovered

  #1816·Buddychek opened Oct 10, 2026·

  ·

• test: gate greedy output across speculative windows and TCP

  #1813·CC-David-CC opened Oct 10, 2026Contributor·

  ·

• build: link CUDA MMQ target to the CUDA driver API

  #1812·CC-David-CC opened Oct 10, 2026Contributor·

  ·

• Community benchmark: RX 9070 16 GB (gfx1201) on Windows 11, IQ3_XXS, engine 0.1.41

  #1811·spitzerd opened Oct 10, 2026·

  ·

• feat: persist conversation checkpoints with RAM and disk restore

  #1810·CC-David-CC opened Oct 10, 2026Contributor·

  ·