All pull requests

Pull requests

• [OMNIML-6340] Add GPTQ-initialized GSQ refinement for native IQ formats

  #3263·ajrasane opened Nov 8, 2026Contributor·

  ·
  4

• Use calibrated per-block amax when compressing static NVFP4 weights

  #3260·udsy19 opened Nov 7, 2026·

  ·
  12

• Use calibrated amax when compressing INT4 and INT8 weights

  #3259·udsy19 opened Nov 7, 2026·

  ·
  7

• Fix AWQ clipping reference with deferred weight scaling

  #3258·taking-lying-flat opened Nov 7, 2026Contributor·

  ·
  4

• Fix example CLI help and QAT dependency installation

  #3257·grzegorz-k-karch opened Nov 7, 2026Contributor·

  ·
  11

• [OMNIML-6062] Dequantize block-FP8 weights that load in the model dtype

  #3256·shengliangxu opened Nov 7, 2026Collaborator·

  ·
  4

• Preserve AWQ clipping thresholds during AutoQuantize folding

  #3255·taking-lying-flat opened Nov 7, 2026Contributor·

  ·
  2

• Export NVFP4 weights after LSQ scale learning

  #3253·rohansjoshi opened Nov 6, 2026Contributor·

  ·
  5

• Export the Qwen3.5 MTP head's fc and norms from Megatron-Core

  #3252·rohansjoshi opened Nov 6, 2026Contributor·

  ·
  3

• Skip the calibration entry when llm_eval overrides the sparse attention backend

  #3251·rohansjoshi opened Nov 6, 2026Contributor·

  ·
  4

• [OMNIML-6461] Run AutoQuantize on expert-parallel and replicated models

  #3250·shengliangxu opened Nov 6, 2026Collaborator·

  ·
  6

• Fix FP8 weight export saturation

  #3249·taking-lying-flat opened Nov 6, 2026Contributor·

  ·
  4

• [NVBug 7018024][fix] Load fused-expert MoE checkpoints in init_quantized_weights

  #3248·noeyy-mino opened Nov 6, 2026Contributor·

  ·
  10

• [NVBug 6927434] Fix the Wan VAE Hugging Face export declaring NVFP4 for unquantized Conv3d weights

  #3247·yingguo-trt opened Nov 6, 2026Contributor·

  ·
  9

• Honor configured attention head dimensions in ONNX export

  #3246·sylvesterkaczmarek opened Nov 6, 2026Contributor·

  ·
  1

  2

• Fix init_quantized_weights: dtype="auto" and trust_remote_code for the model build

  #3244·spped2000 opened Nov 6, 2026Contributor·

  ·
  4

• Add compact base checkpoint export for speculative training

  #3243·sragrawal opened Nov 6, 2026Contributor·

  ·
  13

• [2/2] [OMNIML-6340] Support dense Qwen, Mistral and Gemma in GSQ

  #3241·ajrasane opened Nov 5, 2026Contributor·

  ·
  2

• [2/2] [OMNIML-6340] Add distributed GSQ calibration and restart

  #3240·ajrasane opened Nov 5, 2026Contributor·

  ·
  2

• [OMNIML-6340] Add experimental RCO precision selection to AutoQuant

  #3239·ajrasane opened Nov 5, 2026Contributor·

  ·
  3

• [OMNIML-6056] Take the FSDP2 path in hf_ptq when the plan FSDP2-shards the whole model

  #3236·shengliangxu opened Nov 5, 2026Collaborator·

  ·
  6

• fix(speculative): let a resumed DFlash run apply its own training settings

  #3234·h-guo18 opened Nov 5, 2026Contributor·

  ·
  4

• Preserve conversation and template inputs in streaming datasets

  #3231·sragrawal opened Nov 5, 2026Contributor·

  ·
  23

• Fix EAGLE3 training input normalization

  #3228·sragrawal opened Nov 5, 2026Contributor·

  ·
  10

• [OMNIML-6429] [2/2] Move hf_ptq's per-model loading, preview and calibration rules into the model registry

  #3224·shengliangxu opened Nov 5, 2026Collaborator·

  ·
  24