All pull requests

Pull requests

• [OMNIML-6060] [2/3] Load an MoE expert-parallel: shard the experts at load time

  #3136·shengliangxu opened Oct 30, 2026Collaborator·

  ·
  3

• [OMNIML-6076] [2/3] Prepare a model's config for expert parallelism from its spec

  #3135·shengliangxu opened Oct 30, 2026Collaborator·

  ·
  5

• [OMNIML-6076] [1/3] Describe expert parallelism in the model specs

  #3134·shengliangxu opened Oct 30, 2026Collaborator·

  ·
  9

• [OMNIML-6085] [1/3] Add forward= to distributed_forward_loop, fix the NVML probe, and FSDP2-load text-only Qwen3.5 VLMs through load_pretrained

  #3133·shengliangxu opened Oct 30, 2026Collaborator·

  ·
  5

• [OMNIML-6079] [1/2] Fix Kimi-K3's own modeling code for current transformers and fla-core

  #3132·shengliangxu opened Oct 30, 2026Collaborator·

  ·
  11

• [OMNIML-6060] [1/3] Add require_transformers, and build a VLM without its LM head

  #3131·shengliangxu opened Oct 30, 2026Collaborator·

  ·
  19

• [OMNIML-6063] [1/3] Stage and keep patched a checkpoint's own modeling code

  #3130·shengliangxu opened Oct 29, 2026Collaborator·

  ·
  12

• [OMNIML-6358] Fix NVFP4 normalization rounding

  #3129·meenchen opened Oct 29, 2026Contributor·

  ·
  2

• [OMNIML-6081] Fix the FSDP2 export of W4A8 (SequentialQuantizer) models, and refuse compressing them

  #3128·shengliangxu opened Oct 29, 2026Collaborator·

  ·
  28

• Support vision-language models in the FSDP2 PTQ flow

  #3127·sugunav14 opened Oct 29, 2026Contributor·

  ·
  15

• Fix: add missing attrib

  #3126·haoxiz-nvidia opened Oct 29, 2026Contributor·

  ·
  7

• Bump CI Megatron-Bridge to e42042fa8 and test Qwen3.5-VL bridge export with full NVFP4

  #3125·kevalmorabia97 opened Oct 29, 2026Collaborator·

  ·
  2

• Keep GatedDeltaNet beta / alpha gates unquantized in Megatron fake quant

  #3124·kevalmorabia97 opened Oct 29, 2026Collaborator·

  ·
  31

• [OMNIML-6340] Add GSQ weight calibration and RCO AutoQuant allocation

  #3123·ajrasane opened Oct 29, 2026Contributor·

  ·
  3

• [OMNIML-6073] [1/2] Calibrate and fold expert-parallel fused experts on each rank's own experts

  #3122·shengliangxu opened Oct 29, 2026Collaborator·

  ·
  11

• fix(quant): export the constant FP8 amax

  #3121·mxinO opened Oct 29, 2026Contributor·

  ·
  4

• [8/8] Reload source weights for frozen AutoQuant final calibration

  #3120·meenchen opened Oct 29, 2026Contributor·

  ·
  2

• [7/8] Construct the frozen mixture on authenticated source weights

  #3119·meenchen opened Oct 29, 2026Contributor·

  ·
  2

• [6/8] Bind layerwise recovery to the frozen AutoQuant assignment

  #3118·meenchen opened Oct 29, 2026Contributor·

  ·
  2

• [5/8] Dispatch selected algorithms and estimate source activations

  #3117·meenchen opened Oct 29, 2026Contributor·

  ·
  2

• [4/8] Generate bounded AutoQuant launcher tasks from recipes

  #3116·meenchen opened Oct 29, 2026Contributor·

  ·
  2

• [3/8] Bound automatic exports and recover interrupted workspaces

  #3115·meenchen opened Oct 29, 2026Contributor·

  ·
  2

• [2/8] Calibrate and score independent AutoQuant formats

  #3114·meenchen opened Oct 29, 2026Contributor·

  ·
  2

• [1/8] Plan AutoQuant execution from resolved recipes

  #3113·meenchen opened Oct 29, 2026Contributor·

  ·
  2

• Use model head_dim for Puzzletron runtime benchmarks

  #3112·efegokdemir opened Oct 29, 2026·

  ·
  1

  4