All pull requests

Pull requests

• core: read the model's geometry from its GGUF metadata and refuse a shape the kernels are not compiled for

  #2179·gt50 opened Oct 12, 2026·

  ·

• sm_61 (Tesla P40): a wide dequant kernel, expert mode 8 and its matvec table by default (bitwise), the dense GEMMs as cublasSgemm opt-in (7K prompt 360-432 -> 464-474 tok/s)

  #2178·Techannel opened Oct 12, 2026·

  ·

• STRATA_ROUTE_PRIOR_MAP: the route prior per layer (opt-in); off in layers 0-11 keeps the speed with KL 0.141 -> 0.108

  #2176·zsariboga opened Oct 12, 2026·

  ·

• serve: "env_to_user" (opt-in) moves opencode's <env> block into the first user message (new session's first request 11.9 -> 1.1 s)

  #2175·zsariboga opened Oct 12, 2026·

  ·

• pool: STRATA_POOL_UMWAIT - parked workers sleep with UMONITOR/UMWAIT instead of spinning (opt-in, -3.1% ms/window on a 12 GB laptop, bitwise)

  #2174·zsariboga opened Oct 12, 2026·

  ·

• Community benchmark: RTX 5070 Ti 16 GB, Threadripper 7960X, 64 GB RAM (IQ2_XS, IQ3_XXS, IQ3_S)

  #2173·rayuel opened Oct 12, 2026·

  ·

• setup: install the CUDA Toolkit on Debian 12/13 too, not Ubuntu only

  #2172·gabrielpopa opened Oct 12, 2026·

  ·

• serve: fix a race that makes a long prompt read give way to a short request one chunk late, or never; its TTFT -64% where it hit (#656)

  #2170·SuperDistance opened Oct 12, 2026·

  ·

• serve: refuse a port another program holds on 127.0.0.1, and say which program

  #2168·ByteJuggler opened Oct 12, 2026·

  ·

• bench: RTX 5090 + RTX 3090 two-host stage-RPC UD-Q4_K_XL results

  #2166·petria opened Oct 12, 2026·

  ·

• bench: RTX 5090 + Ryzen 7 9800X3D, Qwen3.8-Flash-Next IQ3_XXS @ 131k (Strata 0.1.43.1)

  #2165·kompotex opened Oct 12, 2026·

  ·

• web: runtime i18n layer (English + Chinese), zero markup changes

  #2164·sdjger-xiaoniu opened Oct 12, 2026·

  ·

• bench: 2x RTX 3080 Ti 12 GB community results (Flash-Next Q4_K_M abliterated, 196K ctx, layer split)

  #2163·tokyorave opened Oct 12, 2026·

  ·

• setup: an interrupted engine update is tried again instead of skipped for good (#2159)

  #2161·ByteJuggler opened Oct 12, 2026·

  ·
  1

• Community benchmark: RTX 3090, Qwen3.8-Flash-Next IQ3_S

  #2160·nacimuss opened Oct 12, 2026·

  ·

• Benchmark results for RTX 3080 Ti and Core Ultra 7 270K Plus

  #2158·XEUIPR opened Oct 12, 2026·

  ·

• bench: RTX 4090 community results (flash-next UD-IQ4_XS 204K/131K, Swift 1.5 IQ3_S)

  #2156·Have-Purple opened Oct 12, 2026·

  ·

• serve: answer GET /models/sse (llama.cpp's model status feed)

  #2155·Daniel-Aaron-Bloom opened Oct 12, 2026·

  ·
  1

• docs: the xe expert tier has no SSD fallback, and --spec 0 is refused on native packs (#1796)

  #2153·nekomario28 opened Oct 12, 2026Contributor·

  ·
  1

• hip(RDNA4): a parity-gated fused-GR register-half down kernel for gfx1200/gfx1201 decode

  #2152·bsorensen110 opened Oct 12, 2026Contributor·

  ·
  9

• SYCL: preserve the borrowed default queue during teardown

  #2151·alex4o opened Oct 12, 2026·

  ·

• SYCL: opt in to 256 GRFs for fused hyperconnection up

  #2150·alex4o opened Oct 12, 2026·

  ·

• hip: add a gfx1150 hipBLASLt 100500 tuning table (Radeon 890M / Ryzen AI 9 HX 370)

  #2148·onionbutties opened Oct 12, 2026·

  ·
  1

• setup: opt-in gfx1150 (Radeon 890M / 880M, Strix Point) as a unified-memory AMD card

  #2147·onionbutties opened Oct 12, 2026·

  ·
  1

• sm_75: the v2 prompt attention in one stage, three blocks per SM (bitwise, attention 1.66x; stacked on #2092)

  #2146·rafatxf opened Oct 12, 2026Contributor·

  ·