All pull requests

Pull requests

• Support explicit embedding scaling in offline fake bases

  #3051·mwawrzos opened Oct 27, 2026·

  ·
  4

• Fix offline speculative label alignment

  #3050·mwawrzos opened Oct 27, 2026·

  ·
  2

• Read config.dtype instead of the deprecated config.torch_dtype alias

  #3049·kevalmorabia97 opened Oct 27, 2026Collaborator·

  ·
  6

• Fold NVFP4 static block-scale prep into the kernel

  #3048·namgyu-youn opened Oct 27, 2026Contributor·

  ·
  4

• Fix: trt missing on arm64

  #3046·haoxiz-nvidia opened Oct 27, 2026Contributor·

  ·
  6

• Check max-calibration amax once instead of per batch

  #3045·namgyu-youn opened Oct 27, 2026Contributor·

  ·
  2

• Collect imatrix statistics with one fused float32 reduction

  #3044·cjluo-nv opened Oct 27, 2026Collaborator·

  ·
  5

• Encode each GPU's imatrix weights in its own thread

  #3043·cjluo-nv opened Oct 27, 2026Collaborator·

  ·
  6

• Skip local Hessian's max-calibration forward when no activation needs it

  #3042·cjluo-nv opened Oct 27, 2026Collaborator·

  ·
  3

• Warn when layerwise export ignores save_modelopt_state

  cherry-pick-0.48.0
  #3041·yingguo-trt opened Oct 27, 2026Contributor·

  ·
  1

  5

• Keep EAGLE-3 vocab maps in their own dtype when quantizing drafters

  #3038·marcelkandula opened Oct 26, 2026·

  ·
  2

• Add a fake-quantized weight cache for QAT and QAD

  #3034·kevalmorabia97 opened Oct 26, 2026Collaborator·

  ·
  16

• [OMNIML-6301] Add a shared per-user cache directory and use it for the AutoTune timing cache and RULER

  #3031·shengliangxu opened Oct 26, 2026Collaborator·

  ·
  50

• Add --freeze_router and --train_moe_expert_fc_only to the Megatron-Bridge distill example

  #3029·yueshen2016 opened Oct 26, 2026Contributor·

  ·
  10

• docs: add Simplified Chinese README (README.zh.md)

  #3021·JasonYeYuhe opened Oct 26, 2026·

  ·
  8

• Add a bfloat16 option for the GPTQ and local-Hessian Hessian GEMMs

  #3019·cjluo-nv opened Oct 26, 2026Collaborator·

  ·
  11

• Fix ONNX export for initializers larger than 2 GiB

  #3013·Tracy-123-dev opened Oct 26, 2026·

  ·
  51

• Fake-quantize all HF fused NVFP4 experts in one kernel launch

  #3011·sychen52 opened Oct 25, 2026Contributor·

  ·
  12

• perf(speculative): fuse the DFlash-family loss with the lm_head projection

  #3010·h-guo18 opened Oct 25, 2026Contributor·

  ·
  4

• Skip Qwen3.8's QSA indexer loop during calibration when it keeps every token

  #3008·cjluo-nv opened Oct 25, 2026Collaborator·

  ·
  5

• Document installing flash-linear-attention for linear-attention models

  #3003·cjluo-nv opened Oct 25, 2026Collaborator·

  ·
  5

• Run fused-MoE expert GPTQ updates on all of this process's GPUs

  #3002·cjluo-nv opened Oct 25, 2026Collaborator·

  ·
  3

• perf(speculative): deliver training batches out of order; compile + FlexAttention in the DFlash2 recipe

  #3001·h-guo18 opened Oct 25, 2026Contributor·

  ·
  4

• feat(launcher): prepare required assets before submission

  #2999·ChenhanYu opened Oct 25, 2026Collaborator·

  ·
  6

• Fix _get_generator is_manual flag and seed=0 handling

  waiting for feedback
  #2998·udsy19 opened Oct 24, 2026·

  ·
  3