All pull requests

Pull requests

• Fix llm_sparsity finetune saving a sharded state dict under FSDP2

  #2919·kaix-nv opened Oct 21, 2026Contributor·

  ·
  3

• [OMNIML-3624] Add modelopt.cli.add_model_args for shared command-line arguments

  #2918·shengliangxu opened Oct 21, 2026Collaborator·

  ·
  5

• build(deps): Bump the github-actions group across 2 directories with 4 updates

  dependenciesgithub_actions
  #2916·dependabot[bot] opened Oct 21, 2026Bot·

  ·
  4

• [2/2] docs: GLM-5.3-Flash IQ2_S and IQ2_S + NVFP4 PTQ recipes and QAD results announcement

  #2915·kevalmorabia97 opened Oct 21, 2026Collaborator·

  ·
  13

• [OMNIML-6228] Seed the quantizer tests that compare against the reference at exact rounding ties

  #2914·shengliangxu opened Oct 21, 2026Collaborator·

  ·
  5

• fix(onnx): preserve NVFP4 compute types

  #2913·host452b opened Oct 21, 2026·

  ·
  2

• Megatron-Bridge distillation: packed chat-SFT data prep and per-document attention

  #2912·kevalmorabia97 opened Oct 21, 2026Collaborator·

  ·
  10

• fix(deploy): unblock TensorRT 11 inference

  cherry-pick-0.48.0
  #2908·joe-nv opened Oct 21, 2026Contributor·

  ·
  4

• Speed up CI JIT compiles: Triton cache and arch-limited CUDA wheels

  #2905·shengliangxu opened Oct 21, 2026Collaborator·

  ·
  4

• Fix attention quantizer leakage in diffusers HF exports

  #2904·Zhehao-Hu opened Oct 20, 2026Contributor·

  ·
  7

• [OMNIML-6227] Convert GGML checkpoints to INT8 compressed-tensors for vLLM

  #2903·cjluo-nv opened Oct 20, 2026Collaborator·

  ·
  30

• Fix vanilla DSpark checkpoint exports for vLLM

  #2902·sragrawal opened Oct 20, 2026·

  ·
  5

• docs(qwen3.6): isolation/loss-mask rerun, chat-SFT kernel correction, Omniscience calibration, announcement update

  #2899·kevalmorabia97 opened Oct 20, 2026Collaborator·

  ·
  31

• Add ResInfer residue quantization for NVFP4 and MXFP4

  #2898·ShawRong opened Oct 20, 2026·

  ·
  4

• Fix GPT-OSS final norm reconstruction in FakeBaseModel

  #2897·sragrawal opened Oct 20, 2026·

  ·
  2

• [OMNIML-6104] Enable DeepSeek-V4-Flash IQ2_XXS PTQ and mergeable LoRA QAD

  #2896·hychiang-git opened Oct 20, 2026Contributor·

  ·
  2

• fix: make the VLM pixel cap bind for images, and warn on an empty assistant mask

  #2895·yeyu-nvidia opened Oct 20, 2026Contributor·

  ·
  2

• launcher: default NUM_NODES from SLURM_NNODES so multi-node jobs do not silently run single-node

  #2894·yeyu-nvidia opened Oct 20, 2026Contributor·

  ·
  4

• test

  #2893·grzegorz-k-karch opened Oct 20, 2026Contributor·

  ·
  6

• Skip MTP forward and recomputation when freeze_mtp is enabled

  #2891·wiederse opened Oct 20, 2026·

  ·
  2

• [2/2] Export vLLM fake-quant checkpoints through Megatron-Bridge

  #2889·kevalmorabia97 opened Oct 20, 2026Collaborator·

  ·
  5

• [1/2] Reload versioned quantizer recipes in vLLM fake-quant serving

  #2888·kevalmorabia97 opened Oct 20, 2026Collaborator·

  ·
  16

• examples: add experimental NeMo Gym evaluation for HF QAD

  #2887·danielkorzekwa opened Oct 20, 2026Contributor·

  ·
  11

  

• [OMNIML-5940] Document Hugging Face QAD workflow in the QAD skill

  #2886·danielkorzekwa opened Oct 20, 2026Contributor·

  ·
  6

• [https://nvbugs/6820271] Fix skillevaluator schema findings in monitor and evaluation

  cherry-pick-0.48.0
  #2885·noeyy-mino opened Oct 20, 2026Contributor·

  ·
  2