All pull requests

Pull requests

• Fix ONNX FP8 calibration scale conversion

  #3102·Michael-RDev opened Oct 29, 2026Contributor·

  ·
  5

• [OMNIML-6082] [2/3] Move each tensor of a sharded model whole to one owner rank

  #3101·shengliangxu opened Oct 29, 2026Collaborator·

  ·
  27

• [OMNIML-6082] [1/3] Let a reverse conversion record a merge rule no key matched

  #3100·shengliangxu opened Oct 29, 2026Collaborator·

  ·
  13

• [OMNIML-6081] [1/2] Pack sharded weights in FSDP2 weight-update windows, and NVFP4 DTensor shards

  #3099·shengliangxu opened Oct 28, 2026Collaborator·

  ·
  32

• Document vLLM fakequant layer and backend support

  #3098·kinjalpatel27 opened Oct 28, 2026Contributor·

  ·
  8

• [3/3] Compile and run a scoped calibration plan from algo_cfg

  #3097·Fridah-nv opened Oct 28, 2026Contributor·

  ·
  2

• [2/3] Declare what each calibration algorithm does

  #3096·Fridah-nv opened Oct 28, 2026Contributor·

  ·
  2

• [1/3] Add a write-mask so calibration can be confined to a scope

  #3095·Fridah-nv opened Oct 28, 2026Contributor·

  ·
  18

• Drop version checks for long-unsupported dependency releases

  #3093·kevalmorabia97 opened Oct 28, 2026Collaborator·

  ·
  5

• Bump transformers to >=5.8, diffusers to >=0.37 and drop older-version code paths

  #3092·kevalmorabia97 opened Oct 28, 2026Collaborator·

  ·
  15

• specdec_bench: report timings in AIPerf vocabulary with tail percentiles

  #3087·yeyu-nvidia opened Oct 28, 2026Contributor·

  ·
  7

• Update SPEED-Bench v1.1 preparation script

  #3086·talorabr opened Oct 28, 2026·

  ·
  16

• [4/7] Add atomic calibration storage with generation reclamation

  #3084·meenchen opened Oct 28, 2026Contributor·

  ·
  2

• [2/7] Preserve calibrated FP8 and mixed-format exports

  #3083·meenchen opened Oct 28, 2026Contributor·

  ·
  2

• [1/7] Preserve offloaded weights and layerwise capture state

  #3082·meenchen opened Oct 28, 2026Contributor·

  ·
  4

• feat(speculative): data echoing for serve-bound draft training

  #3081·h-guo18 opened Oct 28, 2026Contributor·

  ·
  4

  

• [3/9] Serve GDN/KDA state QDQ checkpoints in vLLM

  #3078·kaix-nv opened Oct 28, 2026Contributor·3/9·

  ·
  2

• [1/9] Add independent stochastic TensorQuantizer rounding

  #3077·kaix-nv opened Oct 28, 2026Contributor·1/9·

  ·
  4

• fix(speculative): add the Muse Glimmer final norm for streaming EAGLE3/DFlash training

  #3076·aniskumar-nv opened Oct 28, 2026·

  ·
  2

• [OMNIML-6073] [3/3] Calibrate fused experts on grouped_mm and torch_a2a instead of forcing eager

  #3074·shengliangxu opened Oct 27, 2026Collaborator·

  ·
  140

• Unify weight and attention sparsity under mts.sparsify

  #3073·rohansjoshi opened Oct 27, 2026Contributor·

  ·
  14

• Fix LSQ scale gradients in TE / Megatron layers and add scale learning to the Megatron-Bridge distill.py

  #3069·rohansjoshi opened Oct 27, 2026Contributor·

  ·
  14

• [NVBug 6938144] Qualify the extension fallback in the vLLM example docs

  cherry-pick-0.48.0
  #3068·shengliangxu opened Oct 27, 2026Collaborator·

  ·
  6

• [OMNIML-6073] Quantize Nemotron-H's experts under grouped_mm, with their wrapper in its model folder

  #3067·shengliangxu opened Oct 27, 2026Collaborator·

  ·
  70

• fix: retry transient media loads in the VLM collator

  #3062·yeyu-nvidia opened Oct 27, 2026Contributor·

  ·
  9