All pull requests

Pull requests

• Add mergeable LoRA QAD for Megatron Bridge

  #2656·ChenhanYu opened Oct 5, 2026Collaborator·

  ·
  3

• Preserve histogram counts during entropy calibration

  #2655·sylvesterkaczmarek opened Oct 4, 2026·

  ·
  1

  2

• Fix nested distillation forward contexts

  #2653·Michael-RDev opened Oct 3, 2026·

  ·
  2

• testing doctest

  #2652·grzegorz-k-karch opened Oct 3, 2026Contributor·

  ·
  3

  

• Export unquantized FP8Linear layers in the export dtype, not fp32

  #2651·tritsystem opened Oct 3, 2026·

  ·
  2

• Bump transformers support to <5.19 (5.18) and handle 5.15-5.18 breaking changes

  #2649·kevalmorabia97 opened Oct 2, 2026Collaborator·

  ·
  10

• Backfill checkpoint aliases for Kimi-K2.7-Code and the Gemma-4 MoE releases

  #2647·shengliangxu opened Oct 2, 2026Collaborator·

  ·
  3

• docs(eval): use conservative Terminal-Bench concurrency defaults

  #2646·chadvoegele opened Oct 2, 2026Contributor·

  ·
  3

• fix(distill): flatten MFTLoss labels along with the logits

  #2645·SID-6921 opened Oct 2, 2026·

  ·
  8

• Re-key the GLM-5.3-Flash recipe to its BF16 source checkpoint

  #2644·shengliangxu opened Oct 2, 2026Collaborator·

  ·
  3

• [6/6] Add a sparse teacher-policy dumper

  #2640·phicksy opened Oct 2, 2026·

  ·
  4

• [5/6] Add the tvd_ce objective

  #2639·phicksy opened Oct 2, 2026·

  ·
  4

• [4/6] Add the sparse teacher-policy training path

  #2638·phicksy opened Oct 2, 2026·

  ·
  4

• [3/6] Add the external-draft recipe and training entrypoint

  #2637·phicksy opened Oct 2, 2026·

  ·
  4

• [2/6] Add the external draft speculative-decoding mode

  #2636·phicksy opened Oct 2, 2026·

  ·
  4

• [1/6] Let the hidden-state dump skip aux_hidden_states

  #2635·phicksy opened Oct 2, 2026·

  ·
  5

• Add Cosmos3 Edge DFlash training recipe

  #2632·ConstBob opened Oct 1, 2026·

  ·
  7

• [5/6] Preserve source config in HF fakequant export

  #2630·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  2

• [4/6] Reload vLLM fakequant checkpoints automatically

  #2629·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  2

• [3/6] Fold grouped expert weights during fakequant export

  #2628·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  2

• [2/6] Export Megatron fakequant quantizer state and YAML

  #2627·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  12

• [1/6] Add vLLM fakequant CLI wrapper and app entry point

  #2626·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  27

• [6/6] Prevent EP ranks from overwriting Megatron export shards

  #2625·kinjalpatel27 opened Oct 1, 2026Contributor·

  ·
  2

• Clarify Day 0 sub-agent serving ownership

  #2622·chadvoegele opened Oct 1, 2026Contributor·

  ·
  4

• Fix q/k/v quantizer wiring in register_attention_for_kv_quant

  #2618·javierdejesusda opened Oct 1, 2026Contributor·

  ·
  1

  6