All pull requests

Pull requests

• test(megatron_bridge): cover context parallelism in the VLM QAD test

  #2592·kevalmorabia97 opened Sep 29, 2026Collaborator·

  ·
  8

• Make the QAT/QAD guide the central place for concepts, background, and framework selection

  #2590·danielkorzekwa opened Sep 29, 2026Contributor·

  ·
  12

  

• [Speculative Decoding] LiLiCorr: offline/streaming training and an optional calibration loss

  #2589·mrusanovsky opened Sep 29, 2026Contributor·

  ·
  5

• examples/megatron_bridge: distill on chat data straight from HuggingFace/JSONL

  #2587·yueshen2016 opened Sep 29, 2026Contributor·

  ·
  10

• Reorganize Hugging Face PTQ and QAT/QAD examples under examples/huggi…

  #2586·danielkorzekwa opened Sep 29, 2026Contributor·

  ·
  13

  

• fix(tools/mcp): stop launcher subprocesses inheriting server stdin

  #2585·somtri opened Sep 29, 2026·

  ·
  1

  2

• fix(onnx): move a negative DequantizeLinear axis when transposing the weight

  #2583·SID-6921 opened Sep 29, 2026·

  ·
  3

• [OMNIML-5943] Add fixed-MTP mixed-precision AutoQuant recipes

  #2582·meenchen opened Sep 29, 2026Contributor·

  ·
  2

• [OMNIML-5942] Add MTP quantization calibration and export support

  #2581·meenchen opened Sep 29, 2026Contributor·

  ·
  5

• Move model-specific PTQ modeling into modelopt/torch/models [5/5]

  #2580·shengliangxu opened Sep 28, 2026Collaborator·

  ·
  2

• Move model-specific PTQ modeling into modelopt/torch/models [4/5]

  #2579·shengliangxu opened Sep 28, 2026Collaborator·

  ·
  2

• Move model-specific PTQ modeling into modelopt/torch/models [3/5]

  #2578·shengliangxu opened Sep 28, 2026Collaborator·

  ·
  2

• Move model-specific PTQ modeling into modelopt/torch/models [2/5]

  #2577·shengliangxu opened Sep 28, 2026Collaborator·

  ·
  2

• Move model-specific PTQ modeling into modelopt/torch/models [1/5]

  #2576·shengliangxu opened Sep 28, 2026Collaborator·

  ·
  3

• fix(autocast): do not lose magnitude or sign when narrowing initializers

  #2575·SID-6921 opened Sep 28, 2026·

  ·
  11

• docs: add saturation-aware AutoQuantize blog

  #2574·joshua-hill opened Sep 28, 2026Contributor·

  ·
  7

• Precompile Mamba NAS kernels before functional checks

  #2572·kaix-nv opened Sep 28, 2026Contributor·

  ·
  3

• Fix multi-turn synthetic generation and mark incomplete outputs

  #2571·sychen52 opened Sep 28, 2026Contributor·

  ·
  5

• Deprecate TensorRT-LLM model_type in favor of Hugging Face model_type

  #2569·shengliangxu opened Sep 28, 2026Collaborator·

  ·
  13

• feat(eval): add Tau3-Banking gym task; supersede Tau2-Bench Telecom in the AA suite

  #2568·cjluo-nv opened Sep 28, 2026Collaborator·

  ·
  15

• fix(onnx): keep the converted weight's type in step with its zero point

  #2567·SID-6921 opened Sep 28, 2026·

  ·
  7

• [Reference, not for merge] Move model-specific PTQ modeling into modelopt/torch/models

  #2566·shengliangxu opened Sep 28, 2026Collaborator·

  ·
  4

• [chore]: weekly bump of uv.lock on main (2026-09-28)

  #2564·github-actions[bot] opened Sep 28, 2026Bot·

  ·
  2

• [3/6] Fused Triton GDN/KDA decode QAT

  #2562·kaix-nv opened Sep 28, 2026Contributor·3/6·

  ·
  3

• Improve Puzzletron scoring progress and runtime guidance

  #2560·tejasng053 opened Sep 27, 2026·

  ·
  2