All pull requests

Pull requests

• setup: prefer the cmake/ninja in venv over one on regular PATH

  #2517·zacaj opened Oct 15, 2026·

  ·

• setup: Build on ROCm 10 (TheRock Packaging)

  #2516·brookpatten opened Oct 15, 2026·

  ·

• setup: pin -DSTRATA_Q6K_EXPERTS=ON in the CUDA engine build (stale CMake cache silently drops the Q6_K expert kernels)

  #2515·p3lusa opened Oct 15, 2026·

  ·

• HIP: decode QSA top-k on the register kernel at 262K capacity

  #2514·rindi-ismail opened Oct 14, 2026·

  ·

• prefill: STRATA_HIPBLASLT_TUNING may list one table per architecture

  #2513·rindi-ismail opened Oct 14, 2026·

  ·

• Expose batch MTP draft counts and cache restore payload bytes

  #2511·hotschmoe opened Oct 14, 2026·

  ·

• serve: durable on-disk conversation tier, per-stage segments, prefix restore ladder and Windows chunked large pages

  #2510·shahrokhzargarpour opened Oct 14, 2026·

  ·

• [Needs Testing on Multi-GPU, Nvidia, AMD, Intel, MacOS] Web Engine tab: GPU switches for the user's vendor + in-page restart/save

  #2509·bsorensen110 opened Oct 14, 2026Contributor·

  ·
  23

• Community benchmark: Mac Studio M5 Max 64 GB, IQ3_XXS and UD-Q4_K_XL streamed from the SSD

  #2506·joesinvestments opened Oct 14, 2026·

  ·

• setup: a shard with no tensors is not "short" (UD-Q5_K_XL 00001-of-00006 stopped setup on a correct download)

  #2505·joesinvestments opened Oct 14, 2026·

  ·

• macOS: read available memory for the conversation memory guard

  #2503·joesinvestments opened Oct 14, 2026·

  ·

• resident: fix budget clamp double-counting the pre-claimed arena

  #2501·G0-0000 opened Oct 14, 2026·

  ·

• serve: an engine program that cannot be started is a 503, not a dropped connection

  #2497·devYRPauli opened Oct 14, 2026·

  ·
  1

• sycl: images on the SYCL port (CPU encoder, --vision cpu)

  #2495·skipzx-lab opened Oct 14, 2026·

  ·

• Adding new model/{model id} endpoint to allow use of Strata with Visual Studio

  #2494·MajMcCloud opened Oct 14, 2026·

  ·

• bench: community report - RTX 5060 Ti 16 GB, Coder IQ1_M, 204800-token context (0.1.44.1)

  #2493·Nilabh-code opened Oct 14, 2026·

  ·

• gfx12: empty workgroups past the full-die fill stall in the decode matvecs (grid_pad)

  #2492·i-Vladimir-i opened Oct 14, 2026Contributor·

  ·

• fix(setup): don't require tensor-data alignment for GGUF shards without …

  #2491·OriginalMadman opened Oct 14, 2026·

  ·
  1

• setup: Alchemist on xe also gets the Level Zero v2 adapter (UR_LOADER_USE_LEVEL_ZERO_V2=1; A770 decode +7%, prompt +3%)

  #2490·aslater3 opened Oct 14, 2026Contributor·

  ·

• kernel benches: host-timed (graph + barrier-event profiling hung on an A770 under xe); native_grouped_parity --bench2

  #2489·aslater3 opened Oct 14, 2026Contributor·

  ·

• sycl/A770: faster hyper-connection read on Arc A-series (gr_bench 88/123/151 -> 62/83/100 us at 1/4/6 tokens)

  #2488·aslater3 opened Oct 14, 2026Contributor·

  ·

• sycl/A770: large GRF for the grouped expert kernels (bit-identical; Alchemist default)

  #2487·aslater3 opened Oct 14, 2026Contributor·

  ·

• sycl: Alchemist expert cache in segments, so borrowed prompt GEMM operands are not staged (A770 prompt +8-12%, output identical)

  #2486·aslater3 opened Oct 14, 2026Contributor·

  ·

• sycl: pin the RAM-resident expert copy in 2 GiB chunks when one pinned allocation is refused (Arc A770; prompt DMA for every expert, host staging 3.3 s -> 0.16 s)

  #2485·aslater3 opened Oct 14, 2026Contributor·

  ·

• Community benchmark: 2x RTX PRO 4500 Blackwell, Swift IQ3_XXS, engine 0.1.44.1 (official suite, --long)

  #2484·qni-live opened Oct 14, 2026Contributor·

  ·