All pull requests

Pull requests

• Add GLM-5.3-Flash and GLM-5.2 support to Megatron-Bridge PTQ and HF export

  #2539·kevalmorabia97 opened Sep 23, 2026Collaborator·

  ·
  4

• feat(quantization): run a calibration plan, one mode per stage [4/4]

  #2537·Fridah-nv opened Sep 23, 2026Contributor·4/4·

  ·
  2

• feat(quantization): compile algo_cfg into a validated calibration plan [3/4]

  #2536·Fridah-nv opened Sep 23, 2026Contributor·3/4·

  ·
  4

• feat(quantization): a write-mask so calibration can be confined to a scope [2/4]

  #2535·Fridah-nv opened Sep 23, 2026Contributor·2/4·

  ·
  4

• feat(quantization): declare per-algorithm calibration capabilities [1/4]

  #2534·Fridah-nv opened Sep 23, 2026Contributor·1/4·

  ·
  16

• feat(distill): add kd_loss_weight to KDTrainer for CE+KD loss blending

  #2527·TheSabari07 opened Sep 23, 2026Contributor·

  ·
  2

  4

• fix(speculative): make streaming hidden-state capture work on vLLM 0.29

  #2522·h-guo18 opened Sep 23, 2026Contributor·

  ·
  6

• [2/4] GDN/KDA decode QAT with INT8 recurrent state

  #2519·kaix-nv opened Sep 23, 2026Contributor·2/4·

  ·
  3

• Fix vLLM compilation guard for models without marker

  #2518·kinjalpatel27 opened Sep 23, 2026Contributor·

  ·
  11

• [OMNIML-5899] Export Q8_0 checkpoints and add recipes

  #2517·hychiang-git opened Sep 22, 2026Contributor·

  ·
  4

• [OMNIML-5899] Add Q8_0 quantization codec and backend

  #2516·hychiang-git opened Sep 22, 2026Contributor·

  ·
  5

• [OMNIML-5899] Add Q8_0 CUDA packing kernel

  #2515·hychiang-git opened Sep 22, 2026Contributor·

  ·
  5

• Track every Megatron-Bridge script with MLFlow

  #2514·kevalmorabia97 opened Sep 22, 2026Collaborator·

  ·
  118

• Add the IQ1_M weight-only quantization format

  #2513·cjluo-nv opened Sep 22, 2026Collaborator·

  ·
  5

• Add the IQ2_S weight-only quantization format

  #2512·cjluo-nv opened Sep 22, 2026Collaborator·

  ·
  3

• Use HF converter.rename_source_key when available to reverse naming conversion

  #2508·meenchen opened Sep 22, 2026Contributor·

  ·
  20

• [4/4] Experimental GDN/KDA approximate-inverse QAT

  #2507·kaix-nv opened Sep 22, 2026Contributor·4/4·

  ·
  3

• [3/4] GDN/KDA prefill operand and arithmetic QAT

  #2503·kaix-nv opened Sep 22, 2026Contributor·3/4·

  ·
  3

• Fix grouped expert quantizer checkpoint replicas

  #2500·hychiang-git opened Sep 22, 2026Contributor·

  ·
  30

• feat(distill): support padded vocabulary tails

  #2498·efegokdemir opened Sep 22, 2026·

  ·
  1

  4

• [1/4] GDN state/W QAT foundation

  #2497·kaix-nv opened Sep 22, 2026Contributor·1/4·

  ·
  4

• specdec_bench: forward kv_cache_dtype to the vLLM engine

  #2496·yeyu-nvidia opened Sep 22, 2026Contributor·

  ·
  8

• Make kernels.common.attention independent of kernels.sparsity

  #2495·sychen52 opened Sep 22, 2026Contributor·

  ·
  3

• Fix GitLab authentication guidance for Enroot

  #2493·chadvoegele opened Sep 21, 2026Contributor·

  ·
  3

• Bound the stale-capture warning to once per capture and name its cause

  #2492·Edwardssss opened Sep 21, 2026·

  ·
  1

  7