All pull requests

Pull requests

• Release 2026-10-24 (i4 b3866eb3)

  #309·yy-code-nv opened Oct 24, 2026Collaborator·

  ·
  1

• Merge LoRA adapters safely in automatic Hugging Face exports

  #308·ramanathan831 opened Oct 23, 2026Contributor·

  ·
  3

• Fix delayed learning-rate backoff after loss-spike rollback

  #307·ramanathan831 opened Oct 23, 2026Contributor·

  ·
  3

• fix(guardrail): harden the blocklist against lookalike, fullwidth and invisible spellings; whitelist phrases; measure coverage

  #306·rwagwani opened Oct 23, 2026Collaborator·

  ·

• fix: make training importable and runnable on native Windows

  #303·arif-ahmed-nv opened Oct 22, 2026·

  ·

• fix(mot): gapped replay on GPUs FlashAttention-4 has no kernels for (sm120)

  #302·hakuturu583 opened Oct 21, 2026·

  ·
  7

• cookbook: add Cosmos3-Edge TensorRT-Edge-LLM reasoner and policy notebooks

  #300·ConstBob opened Oct 21, 2026·

  ·
  11

• Add Cosmos3-Edge to the TensorRT-LLM Reasoner cookbook

  #299·ConstBob opened Oct 21, 2026·

  ·
  13

• cookbook(diffusers): add FP8 quantized checkpoint support

  #298·ConstBob opened Oct 21, 2026·

  ·
  4

• fix: admit SM120 and use FA4 with invalid-tile fix

  #295·junghok-nv opened Oct 21, 2026·

  ·
  4

• Expose batched CFG through the inference config

  #279·sunil-srinivasa opened Oct 5, 2026Contributor·

  ·
  1

• Add Cosmos3-Edge variant of the LIBERO-10 action-policy SFT recipe

  #278·filipemartinsubrobotics opened Oct 2, 2026·

  ·
  2

• feat(training): decouple frozen Nano Reasoner for offline and remote SFT conditioning

  #270·Shn0930 opened Sep 22, 2026·

  ·
  2

• fix(inference): decode reasoner video by streaming, not by frame index

  #201·rwagwani opened Aug 14, 2026Collaborator·

  ·
  1

• Make wandb and lerobot optional across default callbacks and datasets

  #169·james77777778 opened Aug 7, 2026Contributor·

  ·

• feat(vlm): LoRA SFT support for VLM + VideoPhy-2 LoRA recipes

  #155·lfengad opened Aug 4, 2026Collaborator·

  ·

• fix(inference): keep transfer attention dense for single samples

  #133·trungtpham opened Jul 24, 2026Contributor·1 of 4 tasks·

  ·

• feat: add compiled_region option to torch.compile the VAE encoder/decoder

  #120·xbcReal opened Jul 21, 2026·

  ·
  2

• Make SFT metadata filters configurable for short-task datasets

  #115·hans-yang-tps opened Jul 16, 2026Contributor·

  ·

• Report MFU on B300 in VLMTokensPerSec

  #114·main-voice opened Jul 16, 2026Contributor·

  ·

• [WIP] LanceDB dataloader

  #100·AyushExel opened Jul 9, 2026·

  ·

• docs(skills): add Cosmos transfer inference skill

  #74·chrisvoncsefalvay opened Jul 1, 2026Contributor·

  ·
  31

• Single-GPU CPU offloading via reasoner/generator split

  #11·rahul-steiger-nv opened Jun 2, 2026Contributor·

  ·
  9