All pull requests

Pull requests

• feat(parallelism): expert parallelism for the Nemotron-3 Nano MoE layers

  #463·le1nux opened Sep 9, 2026·

  ·

• Feat/quality based upsampling

  #461·rrutmann opened Aug 25, 2026·

  ·

• fix(model): initialize wrapped models, bound trunc_normal_, keep all pipeline stages

  #460·le1nux opened Jul 30, 2026·

  ·

• feat(model): add Nemotron-3 Nano hybrid Mamba-Transformer MoE architecture

  #459·le1nux opened Jul 28, 2026·

  ·
  4

• feat: Add torchtitan-style chunked lm_head cross-entropy loss (ChunkedLMHeadCrossEntropyLoss)

  #458·rrutmann opened Jul 16, 2026·

  ·
  

• Add automatic Downstream Evaluation to Modalities - Simplified Implementation

  #455·CYHSM opened Jun 19, 2026·

  ·
  6

• WIP: Mixture-of-Experts

  #454·rrutmann opened Jun 18, 2026·

  ·

• Add automatic Downstream Evaluation to Modalities

  #450·CYHSM opened Jun 9, 2026·

  ·

• Additional logging from Model definition

  #448·CYHSM opened Jun 2, 2026·

  ·

• Context Parallelism

  #446·rrutmann opened May 19, 2026·

  ·
  7

  

• feat: added logging for chunk switches and initial positions in Combi…

  #442·le1nux opened Apr 17, 2026·

  ·

• Added inter document masking for manual and flash attention.

  #434·BlueCrescent opened Feb 23, 2026·

  ·
  37

• Hf checkpoint conversion for distributed checkpoints

  #424·BlueCrescent opened Nov 28, 2025·

  ·
  18

• Added script for updating old checkpoints and configs.

  #397·BlueCrescent opened Aug 25, 2025·

  ·

  · 100B
  1

  15

  

• Improve inference output and add optional system prompt

  #393·CYHSM opened Aug 18, 2025·

  ·
  3

• Instruction Tuning Improvements

  #380·le1nux opened Jul 3, 2025·

  ·
  2

• Eval modalities2

  #306·BlueCrescent opened Feb 21, 2025·

  ·
  4

• Tokenization at scale

  enhancement
  #287·le1nux opened Jan 2, 2025·

  ·

  · v0.3.2
  

• fix: checkpoint conversion to HF

  #285·flxst opened Dec 16, 2024·

  ·

  · 100B
  1

  1