All pull requests

Pull requests

• vulkan : Strix Halo prefill flash attention (KV contiguize, multi-row coopmat1 FA with V^T)

  ggmltestingVulkan
  #396·amangupta-tether opened Oct 16, 2026·

  ·
  5

• ggml, ggml-cuda : keep reserved allocation plans, launch peer-waiting CUDA splits as one graph, long-row TOP_K by radix

  CUDAggmltesting
  #391·pratiknarola-t opened Oct 15, 2026·2/3·

  ·

• opencl : speed up MTP verify and drafting on Adreno (phase 2)

  documentationggmlmodelOpenCLservertesting
  #390·pratiknarola-t opened Oct 15, 2026·

  ·

• common : restore thread-local log levels with backend reload coverage

  devopsggmltesting
  #389·opaninakuffo opened Oct 14, 2026·1 task·

  ·

• QVAC-26703 [Fabric] Port speech-ggml changes onto fabric-llm one - part 4: Vulkan shaders

  documentationggmlVulkan
  #388·mexxik opened Oct 14, 2026·

  ·
  5

• llama : support stream gaps in DSA cache

  modeltesting
  #383·gianni-cor opened Oct 13, 2026·2/6·

  ·

• llama : support stream gaps in DSV4 cache

  testing
  #381·gianni-cor opened Oct 13, 2026·6/6·

  ·

• llama : support stream gaps in hybrid index cache

  modeltesting
  #380·gianni-cor opened Oct 13, 2026·5/6·

  ·

• llama : support stream gaps in MSA cache

  modeltesting
  #379·gianni-cor opened Oct 13, 2026·4/6·

  ·

• llama : support stream gaps in DSA-SWA cache

  modeltesting
  #378·gianni-cor opened Oct 13, 2026·3/6·

  ·

• opencl : MTP speculative-decoding measurements of #370 on a Galaxy S25 (Adreno 830) and the adb bench script

  documentationggmltesting
  #377·Zbig9000 opened Oct 13, 2026·

  ·

• Metal: add Core ML ANE matrix multiplication path

  Apple Metalggmltesting
  #373·anvarmohi opened Oct 12, 2026·

  ·
  1

• common : accept Qwen3-Coder optional args before required ones

  testing
  #372·jpgaribotti opened Oct 12, 2026·

  ·
  

• Rebase b11259

  AMD ZenDNNandroidApple MetalAscend NPUbuildconversionCUDAdevopsdocumentationggmlHexagonIBM zDNNmodelmtmdOpenCLOpenVINOserver/uiserverSYCLtestingvendorVulkanWebGPU
  #369·makaveli10 opened Oct 9, 2026·

  ·
  3

• llama : decode non-consecutive slots in one pass with split KV streams

  testing
  #355·gianni-cor opened Oct 8, 2026·1/6·

  ·
  11

• glm5-next : fix crashes, garbage and slow layer-split prefill on 3 GPUs

  CUDAggmlmodeltesting
  #353·pratiknarola-t opened Oct 7, 2026·3/3·

  ·
  35

• opencl : multi-column q8_0 GEMV for 2-4 column batches on Adreno

  ggmlOpenCLtesting
  #347·gianni-cor opened Oct 6, 2026·

  ·

• QVAC-26346 ggml: port the speech ggml tree onto fabric

  Apple MetalbuildconversionCUDAdevopsdocumentationexamplesggmlHexagonOpenCLtestingVulkanWebGPU
  #346·mexxik opened Oct 6, 2026·

  ·
  11

• glm5-next: align with upstream GLM5-Next

  conversionmodelmtmdtesting
  #331·amangupta-tether opened Oct 2, 2026·

  ·
  1

• cuda: remove most of the DFlash2 prefill slowdown and speed up verify rounds

  CUDAggmltestingVulkanWebGPU
  #324·pratiknarola-t opened Sep 30, 2026·3/4·

  ·

• xdna : consolidate the decode kernels, fix the per-op path, document the backend

  documentationggmltesting
  #319·maximbibik-itrex opened Sep 29, 2026·9/9·

  ·
  40

• xdna : the prompt on the array from phase-3

  ggmltesting
  #318·maximbibik-itrex opened Sep 29, 2026·8/9·

  ·
  7

• xdna : the fused decode from phase-3: a layer a dispatch, a token a command

  documentationggml
  #317·maximbibik-itrex opened Sep 29, 2026·7/9·

  ·
  11

• vulkan: add opt-in batched staging for host uploads

  documentationggmltestingVulkan
  #314·amangupta-tether opened Sep 28, 2026·

  ·

• QVAC-25534 fix: restore OpenVINO Windows CI

  devopsggmlOpenVINOtesting
  #310·yingying0906 opened Sep 25, 2026·

  ·
  5