Issues

• [feature] Memory-bounded DPO scoring on native HF models

  whitecircle/halo#102 · Filter by author MagellaX opened on Oct 6, 2026

• Integration request from aimlapi.com

  enhancement
  whitecircle/halo#100 · Filter by author hugoaimlapi opened on Oct 5, 2026

• [feature] Cached teacher targets: offline distillation without a resident teacher

  enhancement
  whitecircle/halo#97 · Filter by author Strongich opened on Oct 3, 2026
  1

• [feature] Chunked distillation loss: never materialize full [tokens, V] logits

  enhancement
  whitecircle/halo#96 · Filter by author Strongich opened on Oct 3, 2026
  1

• [feature] Distillation losses: generalized JSD(β), teacher-side reverse KL, top-k + tail bucket

  enhancement
  whitecircle/halo#95 · Filter by author Strongich opened on Oct 3, 2026
  3

  

• [feature] First-class context parallelism for GRPO training

  enhancement
  whitecircle/halo#60 · Filter by author MagellaX opened on Sep 24, 2026
  6

  

• [feature] Enable LoRA for dense HF-native tensor-parallel SFT

  enhancement
  whitecircle/halo#45 · Filter by author MagellaX opened on Sep 23, 2026
  11

  

• [feature] Plans for OPD & GRPO

  enhancement
  whitecircle/halo#38 · Filter by author Strongich opened on Sep 22, 2026
  2

• [feature] Concurrent reads in the lazy safetensors loader

  enhancement
  whitecircle/halo#37 · Filter by author mojsaeed opened on Sep 22, 2026
  1

• [feature] Batch DeepGEMM's per-expert weight quantization loop

  enhancement
  whitecircle/halo#36 · Filter by author mojsaeed opened on Sep 22, 2026
  3

  

• [question] Potential support from PEFT

  keep-openquestion
  whitecircle/halo#35 · Filter by author githubnemo opened on Sep 22, 2026
  4