All pull requests

Pull requests

• reorganize examples + add docs info

  #3385·soyr-redhat opened Oct 24, 2026Collaborator·

  ·
  6

• [Tests] Use torch deterministic algorithms in DDP tests

  ready
  #3381·HDCharles opened Oct 23, 2026Collaborator·

  ·
  5

• Update quant docs

  documentationready
  #3376·bartowski1182 opened Oct 23, 2026Collaborator·

  ·
  9

• Support calibration-dependent quantization of MTP layers

  #3371·Mramaran opened Oct 23, 2026·

  ·
  1

  19

• [Recipe] Keep GPTQ actorder through a recipe.yaml round trip

  #3370·zikrulihsan opened Oct 23, 2026·

  ·
  3

• [GPTQ] Keep an explicit actorder: None from config_groups

  #3367·zikrulihsan opened Oct 23, 2026·

  ·
  1

  7

• [GPTQ] Add token masking support for Hessian accumulation

  #3365·BeichenHuang opened Oct 22, 2026Contributor·

  ·
  5

• [Bugfix] Fix Recipe.from_dict for an explicit group and for no arguments

  bug
  #3360·Saged679 opened Oct 22, 2026·

  ·
  2

  10

• [Bugfix] Keep a recipe reusable across several oneshot calls

  bug
  #3353·Saged679 opened Oct 22, 2026·

  ·
  1

  8

• [AWQ] Add opt-in GQA value-to-output smoothing

  #3347·Edwardssss opened Oct 21, 2026·

  ·
  7

• [Misc] Add install_dsv4_chat_template for DeepSeek-V4

  #3346·kylesayrs opened Oct 21, 2026Collaborator·

  ·
  14

  

• [Modifier] Add MoNe (Mixture of Novices) expert pruning

  ready
  #3343·kylesayrs opened Oct 20, 2026Collaborator·

  ·
  22

  

• Refactor initialiation + Calibration start

  needs-rebaseready
  #3342·Roderick-Wu opened Oct 20, 2026Collaborator·

  ·
  20

• [AWQ] Balance the MoE router in build_hybrid_attention_mappings (Qwen3-Next / Qwen3.5-MoE / Qwen3.6)

  #3338·twu3202 opened Oct 20, 2026·

  ·
  1

  6

• [Bugfix] Keep HIGGS candidate schemes that share a bitwidth name

  #3330·sebxt opened Oct 19, 2026·

  ·
  1

  2

  

• [Bugfix] Accumulate eager MSE observer error in FP32 for large groups

  needs-rebase
  #3326·anikeaty08 opened Oct 18, 2026·

  ·
  1

  10

  

• [Bugfix] Remove oneshot file log handler when the run ends

  #3320·anikeaty08 opened Oct 18, 2026·

  ·
  1

  10

  

• [Bugfix] Clamp num_calibration_samples before distributed partitioning

  #3318·anikeaty08 opened Oct 18, 2026·

  ·
  1

  7

• [Bugfix] Fix model_free_ptq with the mse weight observer

  #3315·anikeaty08 opened Oct 18, 2026·

  ·
  1

  4

• [Bugfix] Fix AWQ grid search for small n_grid values

  #3314·anikeaty08 opened Oct 18, 2026·

  ·
  1

  4

  

• [Bugfix] Target HIGGS experts individually when their schemes differ

  #3302·RaphaelFakhri opened Oct 17, 2026·

  ·
  1

  7

• fix(offset-norm): restore untouched offsets byte for byte

  #3298·evanliuuu opened Oct 16, 2026Contributor·

  ·
  1

  7

• [Feature] Linearize MoE experts inside the sequential onload window

  needs-rebase
  #3297·evanliuuu opened Oct 16, 2026Contributor·

  ·
  1

  11

• docs: add verified AutoAWQ checkpoint conversion walkthrough

  ready
  #3296·xiao-yang25 opened Oct 16, 2026·

  ·
  1

  13

• [WiP] Asynchronous offloading writeback

  #3294·Roderick-Wu opened Oct 15, 2026Collaborator·

  ·
  6