All pull requests

Pull requests

• speculative: add opt-in prompt lookup for DFlash2

  #304·bri-prism opened Oct 3, 2026Collaborator·

  ·

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

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

  ·

• qwen35: allow Vulkan for rows-mode GDN state + Vulkan 890M recipe doc

  documentationmodel
  #299·MrFadiAi opened Sep 30, 2026·

  ·

• vulkan: bf16 SSM state pools + gated_delta_net rows mode (+22% decode on 890M)

  ggmlVulkan
  #298·MrFadiAi opened Sep 30, 2026·

  ·
  3

• 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·

  ·

• cuda: bound the one-column MUL_MAT_ID mat-vec store by the expert row count

  CUDAggmltesting
  #295·professorpalmer opened Sep 29, 2026·

  ·

• feat(metal): add multi-column batch verification kernel for PQ2_0 (M=2..8)

  Apple Metalggml
  #293·nogeonwoo opened Sep 28, 2026·

  ·
  4

• vulkan: support GPUs without hardware DP4A in mul_mat_vec_ptq1_0 and eliminate dequant divergence

  ggmlVulkan
  #291·Tetto-chan opened Sep 28, 2026·

  ·
  2

• arm: NEON + i8mm vec_dot for PTQ1_0

  ggml
  #290·karusrus opened Sep 28, 2026·

  ·

• qwen35: opt-in GDN verify tape that replays kept rows instead of per-row snapshots

  modeltesting
  #289·bri-prism opened Sep 28, 2026Collaborator·

  ·

• cuda+server: q8_0 K/V at the full 262k window on 12 GB (tiered VMM KV cache), MTP drafting at every depth

  CUDAggmlmodelservertesting
  #285·professorpalmer opened Sep 27, 2026·

  ·
  2

• vulkan: add dedicated PQ2_0 mat-vec and BC-250 batch tuning

  ggmltestingVulkan
  #281·renovys opened Sep 26, 2026·

  ·
  3

• metal : static-K32 tensor mul_mm for full-tile Q1_0 products (GGML_METAL_Q1_MM_K32=1)

  Apple Metalggmltesting
  #280·jasontitus opened Sep 26, 2026·

  ·

• metal : opt-in PQ2_0 mat-vec for two columns (GGML_METAL_PQ2_MULTICOL=1)

  Apple Metalggmltesting
  #279·jasontitus opened Sep 26, 2026·

  ·

• metal : PTQ1_0 multi-column mat-vec for 5-8 columns

  Apple Metalggmltesting
  #277·jasontitus opened Sep 26, 2026·

  ·

• qwen35 : in-place recurrent state rows for plain decode (GGML_GDN_ROWS_PLAIN=1)

  model
  #276·jasontitus opened Sep 26, 2026·

  ·
  1

• metal : split graphs into 4 command buffers on iOS

  Apple Metalggml
  #275·jasontitus opened Sep 26, 2026·

  ·

• vulkan: extend PTQ1_0 fast integer-dot path to INTEL_XE1 iGPUs (Arrow Lake-H)

  ggmlVulkan
  #274·rriley-quorum opened Sep 25, 2026·

  ·
  1

  2

• arm: NEON vec_dot for PQ2_0

  ggmltesting
  #265·kiljoy001 opened Sep 24, 2026·

  ·
  6

• ggml: add PQ1_0, the Q1_0 binary codec at group 64

  Apple MetalconversionCUDAexamplesggmltesting
  #253·bri-prism opened Sep 23, 2026Collaborator·

  ·
  1

• vulkan: dedicated mul_mat_vec shader for PTQ1_0 (~3x on the dequant path)

  ggmlVulkan
  #252·Preygle opened Sep 23, 2026·

  ·
  4

• ggml-cpu : move repacked weight rows to the NUMA node of the thread that reads them

  ggml
  #251·lenny76 opened Sep 23, 2026·

  ·
  1

• ggml-cpu : AVX-512 VNNI and AVX2 vec_dot for PTQ1_0 on x86 (4x token generation)

  ggml
  #250·lenny76 opened Sep 23, 2026·

  ·
  6

• ggml-cpu : convert each mul_mat src1 row on a single thread (faster token generation)

  ggml
  #249·lenny76 opened Sep 23, 2026·

  ·
  1