All pull requests

Pull requests

• xdna : fix last-token logits and speed up embedding-model prefill (EmbeddingGemma 2)

  documentationggmltesting
  #484·pratiknarola-t opened Nov 3, 2026·

  ·

• Restore QVAC runner labels after upstream sync

  devopstesting
  #481·mexxik opened Nov 3, 2026·

  ·
  1

• ggml: allocation-free physical buffer sizing with backend guards (temp-11739)

  documentationggmlHexagontesting
  #479·Zbig9000 opened Nov 3, 2026·2 of 4 tasks·

  ·

• ggml-cpu: scope Linux caller affinity to graph execution

  ggmltesting
  #478·ihor-v-d opened Nov 3, 2026·1 task·

  ·

• Restore explicit F32 CUDA matmul precision on temp-11739

  CUDAggmltesting
  #477·mexxik opened Nov 3, 2026·

  ·

• Restore Metal diagonal masking on temp-11739

  Apple Metalggmltesting
  #476·mexxik opened Nov 3, 2026·

  ·

• hexagon: preserve canonical weights across partial REPACK uploads

  ggmlHexagontesting
  #473·Alok-Ranjan23 opened Nov 3, 2026·1 task·

  ·
  1

• ggml-opencl: fix pre-init buffer queries and strided ADD crash

  ggmlOpenCLtesting
  #472·Alok-Ranjan23 opened Nov 3, 2026·

  ·

• mtmd: preserve OCR position tables and correct projector memory

  AMD ZenDNNandroidApple MetalAscend NPUbuildconversionCUDAdevopsdocumentationggmlHexagonIBM zDNNjinja parsermodelmtmdOpenCLOpenVINOserver/uiserverSYCLtestingvendorVulkanWebGPU
  #468·Alok-Ranjan23 opened Nov 3, 2026·

  ·

• Add Hexagon runtime information and optional F32 convolution taps

  documentationggmlHexagontesting
  #465·mexxik opened Nov 2, 2026·

  ·
  1

• metal : few-row MMA mat-mul up to 64 src1 rows, limits per type

  Apple Metalggmltesting
  #462·amangupta-tether opened Nov 2, 2026·

  ·

• coreml : Qwen3.8 FFN sidecar on the Neural Engines, in-graph with load-time export

  Apple Metalbuilddevopsdocumentationexamplesggmlmodeltesting
  #461·amangupta-tether opened Nov 2, 2026·

  ·
  10

• spec : per-size schedulers and per-step p-min for MoE speculative decoding on temp-11259 (port of #31156, #31506)

  modelserver
  #460·amangupta-tether opened Nov 2, 2026·

  ·

• fix(gguf): handle string arrays in gguf_get_arr_data - DoS on valid input

  ggml
  #452·rozzkml opened Oct 28, 2026·

  ·
  1

• fix(server): secure built-in tools - remove exec_shell_command from default, block x-tool-cwd path traversal

  server
  #451·rozzkml opened Oct 28, 2026·

  ·
  1

• fix(server): add SSRF protection for image/audio URL downloads in chat completions

  server
  #450·rozzkml opened Oct 28, 2026·

  ·
  1

• QVAC-26707: refresh op-support docs and speech port closeout (part 8)

  documentation
  #431·mexxik opened Oct 26, 2026·

  ·

• mtmd : add decode limits to mtmd-helper

  mtmdtesting
  #429·dev-nid opened Oct 23, 2026·

  ·

• skills : add engine-bench for the shared Tailscale machines

  documentation
  #424·amangupta-tether opened Oct 23, 2026·

  ·
  1

• infra[notask]: pin setup-node to lts/* with check-latest and setup-bun to latest

  devops
  #423·Proletter opened Oct 22, 2026·

  ·
  1

• hip : experimental W4A4 Q4_0 matmul with q4_0 activations (DOT8 and WMMA IU4)

  CUDAggmltesting
  #412·amangupta-tether opened Oct 21, 2026·

  ·

• 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·

  ·

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

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

  ·
  11

• 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·

  ·