All pull requests

Pull requests

• json-schema-to-grammar : convert empty anyOf / oneOf / type unions instead of emitting an empty rule

  examplestesting
  #333·professorpalmer opened Oct 7, 2026·

  ·

• cuda: GGML_CUDA_FA_PREFILL_F16=N, f16 prefill for quantized K/V from pool memory (decode stays in place)

  CUDAggml
  #330·professorpalmer opened Oct 7, 2026·

  ·

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

  ggmlSYCL
  #328·kiljoy001 opened Oct 7, 2026·

  ·

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

  ggmlSYCL
  #327·kiljoy001 opened Oct 7, 2026·

  ·

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

  ggmlSYCL
  #326·kiljoy001 opened Oct 7, 2026·

  ·

• cuda: take the planar PTQ1_0 layout at one column on Turing as well as Ampere

  CUDAggml
  #325·professorpalmer opened Oct 7, 2026·

  ·

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

  #323·professorpalmer opened Oct 5, 2026·

  ·

• cuda: BATCH_INVARIANT: occupancy-independent FA KV split, PTQ1_0 mat-vec up to 8 columns

  #322·professorpalmer opened Oct 5, 2026·

  ·

• server: --reasoning-effort-allow/-fallback and --reasoning-max-tokens-floor

  server
  #321·professorpalmer opened Oct 5, 2026·

  ·

• speculative: --spec-draft-window and --spec-draft-n-max-tail (MTP drafting at every depth)

  server
  #320·professorpalmer opened Oct 5, 2026·

  ·

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

  #319·professorpalmer opened Oct 5, 2026·

  ·

• cuda : backport GGML_CUDA_FA_QUANTS from upstream (ggml-org/llama.cpp#28079)

  CUDAdocumentationggml
  #317·cheese-cakee opened Oct 5, 2026·

  ·
  1

  3

• llama-quant : refuse PQ2_0/PTQ1_0 output without prism.hadamard metadata

  examples
  #316·cheese-cakee opened Oct 5, 2026·

  ·
  1

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

  CUDAggml
  #314·sb32445 opened Oct 4, 2026·

  ·
  2

• common : take the top-k straight from the logits when top-k starts the chain

  #313·sb32445 opened Oct 4, 2026·

  ·
  1

• server : reuse the buffers of evicted prompt checkpoints

  server
  #312·sb32445 opened Oct 4, 2026·

  ·

• kv-cache : limit seq_rm to the used cell range

  #311·sb32445 opened Oct 4, 2026·

  ·

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

  CUDAggmltesting
  #310·sb32445 opened Oct 4, 2026·

  ·

• cuda: use the transposing concat kernel on all GPUs, not only on GB10

  CUDAggml
  #309·sb32445 opened Oct 4, 2026·

  ·

• cuda: smaller KV tile for the 8-column MMA flash attention config (head size 256)

  CUDAggml
  #308·sb32445 opened Oct 4, 2026·

  ·

• cuda: use the MMA flash attention kernel for GQA above 4 with quantized K/V on Ada

  CUDAggml
  #307·sb32445 opened Oct 4, 2026·

  ·

• cuda: PQ2_0 mat-vec kernel for 3-8 columns on Ada

  CUDAggml
  #306·sb32445 opened Oct 4, 2026·

  ·
  4

• 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 : GET_ROWS -> RESHAPE -> GATED_DELTA_NET case, eager + CUDA graph capture/replay

  testing
  #296·professorpalmer opened Sep 29, 2026·

  ·