All pull requests

Pull requests

• [feat](kt-kernel): add AVX2 RAWINT4 routed-expert LoRA SFT backend

  #2243·losewayy opened Oct 11, 2026·2 tasks done·

  ·

• [perf](kt-kernel): faster llamafile MoE for small token windows

  #2242·05yuki opened Oct 10, 2026Contributor·

  ·

• [perf](kt-kernel): llamafile expert weights on transparent huge pages

  #2241·05yuki opened Oct 10, 2026Contributor·

  ·

• [feat](kt-kernel): x86 llamafile kernels for IQ2_XS, IQ2_S, IQ3_XXS, IQ3_S and IQ4_NL

  #2240·05yuki opened Oct 10, 2026Contributor·

  ·

• [feat](kt-kernel): llamafile MoE with intermediate sizes that are not a multiple of 256

  #2239·05yuki opened Oct 10, 2026Contributor·

  ·

• [fix](kt-kernel): llamafile MoE honours physical_to_logical_map

  #2238·05yuki opened Oct 10, 2026Contributor·

  ·

• [fix](kt-kernel): IQ4_XS kernels (AVX2 and AVX-512) without int16 saturation

  #2237·05yuki opened Oct 10, 2026Contributor·

  ·

• [fix+test](experts): correct index in moe_infer_simple

  #2236·SACHokstack opened Oct 10, 2026·8 tasks done·

  ·

• [perf](kt-kernel): share checkpoint indexes across pending MoE layers

  #2231·ZenAlexa opened Oct 7, 2026·3 tasks done·

  ·
  1

  1

• [fix](kt-kernel): reject AVX2 FP8/RAWINT4 TP splits that break block quantization

  #2229·Dashener2 opened Oct 7, 2026·2 tasks done·

  ·

• [fix](kt-kernel): size per-NUMA thread counters by physical NUMA id

  #2228·Dashener2 opened Oct 7, 2026·2 tasks done·

  ·
  1

  2

• [docs]: fix links to moved injection and Kimi-K2 native tutorials

  #2223·pratikgx opened Oct 1, 2026·1 of 2 tasks·

  ·

• [fix](cli): fill in the model name and path in kt model remove/edit warnings

  #2220·MohammadHijjawi97 opened Sep 30, 2026·2 tasks done·

  ·

• [fix](cli): close the tokenizer warning markup in kt chat

  #2219·MohammadHijjawi97 opened Sep 30, 2026·2 tasks done·

  ·

• [build] Sync sglang-kt submodule (v0.7.0.post4)

  automateddependencies
  #2217·github-actions[bot] opened Sep 29, 2026Bot·

  ·

• Kt hardening

  #2215·SaboniAmine opened Sep 28, 2026·

  ·

• [perf](kt-kernel): tile the activation rows in the AVX2 MXFP4 fast paths

  #2212·sakurahello1 opened Sep 17, 2026·2 tasks done·

  ·
  2

• [fix](cli): keep "1" and "0" as integers in kt config set

  #2201·Dev-next-gen opened Sep 10, 2026·2 tasks done·

  ·

• [perf](sft): accelerate FP8 training with shared repack and decoded tiles

  #2193·yyj6666667 opened Sep 10, 2026Collaborator·

  ·

• [feat](kt-kernel): --kt-mmap-experts-dir to keep CPU expert weights in a file, not anonymous RAM

  #2190·tim-odonnell opened Sep 7, 2026·

  ·

• [fix](kt-kernel): keep CPU expert buffers alive during CUDA graph capture

  #2174·05yuki opened Aug 27, 2026Contributor·1 of 2 tasks·

  ·

• chore: update SGLang to support_xing4_0

  #2168·157933996 opened Aug 19, 2026·

  ·

• fix(kt-kernel): AMXMoEWrapper cpu_save branch crashes on load_experts keys/prefix (#2151)

  #2161·Anai-Guo opened Aug 18, 2026Contributor·

  ·

• [feat]: add AMXINT4 SMART 3-dtype wrapper-conversion MoE routing

  #2155·NukePeace opened Aug 14, 2026·1 of 2 tasks·

  ·
  12

• [WIP][feat](kt-kernel): Add Ascend NPU CPU-MoE backend + llama.cpp b3173 patch support

  #2153·Nikonz opened Aug 13, 2026·2 tasks done·

  ·
  1