All pull requests

Pull requests

• cuda: scale PTQ1_0 MMQ results once per 128-weight block

  CUDAggml
  #411·sb32445 opened Oct 27, 2026·

  ·

• cuda: use ncols2 = 2 in the MMA flash attention for prompt batches at GQA ratio 6

  CUDAggmltesting
  #410·sb32445 opened Oct 27, 2026·

  ·

• cuda: harden the FWHT q8_1 fusions (output-flagged views, packed-output aliasing, GLU tests)

  CUDAggmltesting
  #409·bri-prism opened Oct 26, 2026Collaborator·

  ·

• cuda: keep softcap and int32-overflowing shapes off the D=72 pad path (follow-up to #390)

  CUDAggml
  #408·bri-prism opened Oct 26, 2026Collaborator·

  ·

• qwen35: use GDN state rows for safe single-sequence plain decode

  Apple Metalggmlmodel
  #407·bri-prism opened Oct 26, 2026Collaborator·

  ·

• metal: skip empty asynchronous tensor uploads

  Apple Metalggml
  #406·bri-prism opened Oct 26, 2026Collaborator·

  ·
  1

• ci: add gfx1035 and gfx1103 to the Prism ROCm/HIP release targets

  devops
  #405·bri-prism opened Oct 26, 2026Collaborator·

  ·

• a ternary model needs a ternary drafter

  #404·Cyronius opened Oct 26, 2026·

  ·

• sycl: allocate the PTQ1_0/PQ2_0 and FA scratch buffers with ggml_sycl_malloc_device

  ggmlSYCL
  #384·Torchit1 opened Oct 24, 2026·

  ·

• cuda: opt-in one-scale-per-128 activations for PQ2_0 prefill (GGML_CUDA_PQ2_D1), +4-24% on H200, not bit-exact

  CUDAggml
  #375·bri-prism opened Oct 23, 2026Collaborator·4 of 5 tasks·

  ·

• model: skip prism.hadamard rotation for tensors that were not loaded

  testing
  #372·Torchit1 opened Oct 20, 2026·

  ·
  1

• Merge upstream master into prism (2026-10-11)

  androidApple MetalAscend NPUbuildconversionCUDAdevopsdocumentationggmlHexagonIBM zDNNjinja parsermodelmtmdnixOpenCLOpenVINOserver/uiserverSYCLtestingvendorVulkanWebGPU
  #352·bri-prism opened Oct 11, 2026Collaborator·

  ·
  9

• docs: cover the pip SDK route and runtime pitfalls for Windows HIP builds

  documentation
  #344·Thatgfsj opened Oct 10, 2026·

  ·
  1

• server: --checkpoint-every-nt N, context checkpoints inside a long message (off by default)

  server
  #338·professorpalmer opened Oct 9, 2026·

  ·
  2

• sycl: fuse the PQ2_0/PTQ1_0 XMX mat-muls with their neighbours (depends on #327)

  ggmlSYCL
  #328·kiljoy001 opened Oct 7, 2026·

  ·
  1

• sycl: PQ2_0/PTQ1_0 XMX kernels from TernSYCL, PTQ1_0 kept at 1.75 bpw

  ggmlSYCL
  #327·kiljoy001 opened Oct 7, 2026·

  ·
  1

• sycl: faster concat, get_rows and few-row f16/bf16 mat-vec for short rows

  ggmlSYCL
  #326·kiljoy001 opened Oct 7, 2026·

  ·
  1

• cuda: GGML_CUDA_OP_TIMING=1 per-node GPU time breakdown

  CUDAggml
  #323·professorpalmer opened Oct 5, 2026·

  ·
  2

• speculative: --spec-draft-window (MTP drafting at every depth)

  server
  #320·professorpalmer opened Oct 5, 2026·

  ·
  8

• llama/cuda: tiered KV cache (--kv-vram-cells): VRAM head, pinned-host tail in one VMM range

  CUDAggml
  #319·professorpalmer opened Oct 5, 2026·

  ·
  2

• cuda: prefetch the next PTQ1_0 mat-vec's weights into L2 from the last CTAs

  CUDAggml
  #314·sb32445 opened Oct 4, 2026·

  ·
  3

• cuda: gate (SwiGLU) fused PTQ1_0 mat-vec for 2-4 columns

  CUDAggmltesting
  #310·sb32445 opened Oct 4, 2026·

  ·
  1

• fix(cuda/msvc): add CUDA 13.4 CUB fixes, Win32 CreateFileW, and graceful tensor name fallback

  Apple MetalconversionCUDAdevopsdocumentationexamplesggmltestingvendor
  #300·vrwallace opened Sep 30, 2026·

  ·

• metal: opt-in Metal 4.1 direct Q2 operand experiment

  Apple Metaldocumentationggml
  #297·bri-prism opened Sep 30, 2026Collaborator·

  ·

• tests: MUL_MAT_ID cases for the one-column mat-vec store (small K, several expert rows per block)

  CUDAggmltesting
  #295·professorpalmer opened Sep 29, 2026·

  ·
  2