All pull requests

Pull requests

• [4/4] Add a draft-loss weight so a co-training control arm is expressible

  #2827·yeyu-nvidia opened Oct 15, 2026Contributor·

  ·
  10

• [3/3] Add a packed pretraining data mode for speculative-decoding training

  #2826·yeyu-nvidia opened Oct 15, 2026Contributor·

  ·
  10

• [NVBug 6779760] Test NVFP4 export of CPU-offloaded models

  #2825·cjluo-nv opened Oct 15, 2026Collaborator·

  ·
  2

• Add an optional language-modelling anchor for a trainable DFlash base

  #2824·yeyu-nvidia opened Oct 15, 2026Contributor·

  ·
  7

• Speed up CPU unit tests to stay under the 20 min CI limit

  cherry-pick-0.48.0
  #2823·kevalmorabia97 opened Oct 15, 2026Collaborator·

  ·
  20

• [6914448] Keep Qwen vision blocks in high precision for vLLM

  cherry-pick-0.48.0
  #2822·ajrasane opened Oct 15, 2026Contributor·

  ·
  4

• [OMNIML-6149] Merge the input amax of fused experts that share weights without a declared tie

  #2821·shengliangxu opened Oct 15, 2026Collaborator·

  ·
  5

• Fix scale-aware PTQ for native FP8 experts

  cherry-pick-0.48.0
  #2820·chadvoegele opened Oct 15, 2026Contributor·

  ·
  13

• Fix MiniMax M2.7 expert selectors and source dtype documentation

  #2817·chadvoegele opened Oct 15, 2026Contributor·

  ·
  5

• Fix Windows ONNX external data corruption by upgrading ONNX

  cherry-pick-0.48.0
  #2816·anmaurya001 opened Oct 15, 2026·

  ·
  8

• [6933164] fix: preserve Diffusers pipeline settings in component exports

  cherry-pick-0.48.0
  #2813·yingguo-trt opened Oct 15, 2026Contributor·

  ·
  7

• Upgrade ONNX Runtime 1.30 and CUDA 13 for Linux x86-64

  cherry-pick-0.48.0
  #2809·ajrasane opened Oct 14, 2026Contributor·

  ·
  12

• [6918844] Fix LSQ scale floor

  cherry-pick-0.48.0
  #2808·rohansjoshi opened Oct 14, 2026Contributor·

  ·
  5

• Fix fused-experts export for W4A8 AWQ (SequentialQuantizer) weight quantizers

  cherry-pick-0.48.0
  #2807·Fridah-nv opened Oct 14, 2026Contributor·

  ·
  16

• [4/8] Add dynamic NVFP4 recurrent-state QAT

  #2805·kaix-nv opened Oct 14, 2026Contributor·3/7·

  ·
  4

• [6915422] Fix TensorRT 11 inference shape validation

  cherry-pick-0.48.0
  #2798·ajrasane opened Oct 14, 2026Contributor·

  ·
  11

• Let the multimodal generator enable thinking and fit its output budget

  #2797·yeyu-nvidia opened Oct 14, 2026Contributor·

  ·
  7

• [6915426] Fix ONNX AutoTune benchmark initialization on TensorRT 11

  cherry-pick-0.48.0
  #2795·ajrasane opened Oct 14, 2026Contributor·

  ·
  7

• [5/8] ReplaySSM state QAT and two-precision GDN

  #2788·kaix-nv opened Oct 14, 2026Contributor·4/7·

  ·
  4

• Make freeze_mtp an explicit override of the automatic MTP freeze

  #2784·yueshen2016 opened Oct 13, 2026Contributor·

  ·
  6

• [OMNIML-6060] [12/21] Load an expert-parallel plan through load_pretrained

  #2780·shengliangxu opened Oct 13, 2026Collaborator·13/22·

  ·
  9

• [OMNIML-6054] [6/21] Route load_pretrained by the distribution plan

  #2779·shengliangxu opened Oct 13, 2026Collaborator·7/22·

  ·
  13

• [2/2] Unified export: quantized-weight packing and state transforms

  cherry-pick-0.48.0export/deploy
  #2778·mxinO opened Oct 13, 2026Contributor·

  ·
  11

• [1/2] Unified export: quantized-weight state capture

  cherry-pick-0.48.0export/deploy
  #2777·mxinO opened Oct 13, 2026Contributor·

  ·
  8

• Sync GPTQ Hessians across data-parallel ranks

  cherry-pick-0.48.0
  #2773·kevalmorabia97 opened Oct 13, 2026Collaborator·

  ·
  16