All pull requests

Pull requests

• mtp_pack: Q2_0 scale search over 0.15 to 1.0 x max|w|, either sign

  #1934·gputier opened Oct 11, 2026Contributor·

  ·

• serve: aggregate batch request timings

  #1933·constantindjonkam opened Oct 11, 2026Contributor·

  ·

• fix(hip): add opt-in stability profile for reset reports

  #1931·CC-David-CC opened Oct 11, 2026Contributor·

  ·

• bench: community report, RX 6900 XT (gfx1030) x1 and x2 on 0.1.42 - Unsloth UD-Q4_K_XL beside GSQ-RCO IQ3_S, stock and with the gfx103x PRs

  #1927·Vop3p opened Oct 11, 2026Contributor·

  ·

• Add optional GPU electricity cost display to Power card

  #1925·hash13 opened Oct 11, 2026Contributor·

  ·

• docs: CONFIGURATION.md reference documentation for all flags and environment variables

  #1924·SriRamanujam opened Oct 11, 2026·

  ·

• Community benchmark: 5070 Ti + 2x 5060 Ti (one on a USB4 eGPU), UD-IQ4_XS, engine 0.1.42, Ryzen 7 7700X, 46 GiB RAM in an LXC

  #1921·chiangww opened Oct 11, 2026·

  ·

• OpenAI logprobs on /v1/chat/completions for the CUDA/HIP engine

  #1916·gt50 opened Oct 11, 2026·

  ·

• bench: community report, RTX 5080 Laptop 16 GB, IQ3_XXS, engine 0.1.42

  #1914·markusrenneronline opened Oct 11, 2026·

  ·

• --pool-affinity measure: on a hybrid CPU, time the all and p-cores pool layouts at start and take the faster (opt-in)

  #1912·Xiaomi2008 opened Oct 11, 2026·

  ·

• Expert cache: STRATA_VRAM_FREE=device reads the whole card's free VRAM through NVML (opt-in, Windows + NVIDIA)

  #1911·Xiaomi2008 opened Oct 11, 2026·

  ·

• kernels: process all groups in fused GU and LDS down launches

  #1909·W1nge opened Oct 11, 2026Contributor·

  ·
  1

• arena: pin by need (STRATA_PIN_BY_NEED=1) and a PCIe share floor (STRATA_PCIE_MIN1=1), +5-7 % prompt read on Windows

  #1905·RubenTorney opened Oct 11, 2026·

  ·

• adaptive tier: --adapt-async 1 in the arena mode (one GPU, copy-in only), x1.07-1.11 cold decode on an RTX 3090 + Ryzen 5 5600X

  #1904·RubenTorney opened Oct 11, 2026·

  ·

• test: synchronize the keepawake service fixture

  #1901·midhatn opened Oct 10, 2026Contributor·

  ·

• serve: the thinking budget's continuation keeps the tokens emitted after the STOP, so it is not read again

  #1899·Vokturz opened Oct 10, 2026·

  ·

• setup: build the HIP engine with the K-quant MMQ prompt kernels (UD-Q4_K_XL's experts on RDNA2, 633 -> 1,036 tok/s)

  #1897·elightcap opened Oct 10, 2026·

  ·
  1

• resident RAM mode: a budget keeps the prompt path's lend region when it has room (#1683)

  #1896·elightcap opened Oct 10, 2026·

  ·
  1

• chat: syntax highlighting for fenced code blocks in streamed output

  #1895·MegaRushX opened Oct 10, 2026·

  ·

• bench: community report, RX 6900 XT (gfx1030): STRATA_ROUTE_TAIL_SKIP=7 on HIP, decode x1.24 (Unsloth UD-Q4_K_XL, 0.1.42)

  #1894·elightcap opened Oct 10, 2026·

  ·

• docs: the three things 0.1.42 changed that no page told

  #1893·Buddychek opened Oct 10, 2026·

  ·

• Non-finite logits (#879): a guard with self-healing

  #1892·lask3802 opened Oct 10, 2026·

  ·

• logit_bias: normalize a long list once per request (65,929 entries: +175 ms -> +24 ms)

  #1891·qni-live opened Oct 10, 2026Contributor·

  ·

• sycl: Windows/OpenCL support for the merged port, and a live free-VRAM figure for the expert cache (#1549)

  #1890·demetree opened Oct 10, 2026Contributor·

  ·

• Verify window: the mixer's input-only work on graph branches (STRATA_SIDE_MIXER=1/2, opt-in)

  #1889·KarlGrier opened Oct 10, 2026Contributor·

  ·