All pull requests

Pull requests

• Turing (sm_75): prefill attention on a dequantized window, fa75 kernel for head_dim 256

  #439·rafatxf opened Oct 2, 2026·

  ·

• GatedResidual: EXL3_GR_MIX_TILED=0 drops the duplicate projection tables (fp32-accurate fallback on the decode tables)

  #430·Vemloir opened Oct 1, 2026·

  ·
  1

• DFlash2: speculative (rejection) sampling for stochastic samplers

  #423·rafatxf opened Sep 30, 2026·

  ·

• perf: add opt-in QSA prefill and bounded recurrent checkpoint staging

  #422·PeisongXiao opened Sep 29, 2026·

  ·

• Add experimental Ling-3.0-Flash (BailingMoeV3) support

  #419·everson opened Sep 28, 2026·

  ·

• feat: embed_tokens trellis quantization

  #410·HermiG opened Sep 27, 2026·

  ·
  2

• AVX2 MoE GEMV: group-2 swizzled trellis layout with a band-2 kernel (+4-6% decode)

  #408·treo opened Sep 26, 2026Contributor·

  ·

• Volta (sm_70) support: GEMV decode kernels, arch gating, reference tests

  #404·Frenchy2k1 opened Sep 24, 2026·

  ·

• feat(yaqa): add opt-in recursive collection

  #395·silvanshade opened Sep 22, 2026·

  ·

• SM120 TMA based paged prefill flash attention for D128/256/512

  #394·stevelikesrhino opened Sep 22, 2026·

  ·

• CPU MoE: streamed single-row decode from the pinned arena (EXL3_MOE_STREAM_DECODE)

  #392·lmganon16 opened Sep 21, 2026·

  ·
  11

• Loader: Use dedicated NVML headroom on Windows WDDM

  #389·DenysAshikhin opened Sep 21, 2026Contributor·

  ·

• Add DFlash2 draft support: grouped dynamic convolution + candidate selector, q-aware verify

  #379·am423 opened Sep 16, 2026·

  ·
  2

• Port recent kernel additions for HIP builds

  #367·sdougbrown opened Sep 12, 2026Contributor·

  ·
  1

• Add a PyTorch fallback layer for ROCm-excluded extension functions

  #366·sdougbrown opened Sep 12, 2026Contributor·

  ·
  1

• Make the extension build and load under ROCm/HIP

  #365·sdougbrown opened Sep 12, 2026Contributor·

  ·
  1

• Build the CUDA extension as C++20

  #360·sdougbrown opened Sep 12, 2026Contributor·

  ·
  2

• Add LFM2.5-VL architecture support

  #347·WatchDG opened Sep 8, 2026Contributor·

  ·

• Layer-split forward: keep the current CUDA device on the module's device

  #337·creslinux opened Sep 5, 2026Contributor·

  ·
  2

• Precomputed MoE expert-placement profiles for CPU offload

  #315·anoane opened Sep 1, 2026·

  ·
  9

• Speed up Qwen MTP decoding by 22% with a selected vocabulary head

  #303·keldenl opened Aug 22, 2026·

  ·
  4

• PoC: per-layer split or fused-uniform QKV topology

  #299·malaiwah opened Aug 20, 2026Contributor·

  ·
  3

• exl3_mgemm: fix out-of-bounds slot list, and allow multi-token reduction with expert-range filtering

  #290·arbi-dev opened Aug 17, 2026·

  ·

• MoE: Add fused additive EXL3 kernels

  #284·malaiwah opened Aug 13, 2026Contributor·

  ·
  1

• Rocm gfx1100 integration

  #283·Calandracas606 opened Aug 13, 2026·

  ·