🧰 TorchTitan

The items that name it 5

• Amazon’s Trainium Takes a Strike at the Training King
  article2026-10-23

  Open-source pre-training framework. Used to train Qwen3-32B and gpt-oss on Trainium.

• [AINews] High Taste Testers Ep. 1
  article2026-10-17It is written here as TorchTitan (Zephon).

  Resuming runs on 8, 32 or 64 GPUs from a 16-GPU checkpoint with SHA-256-identical global batches.

• [AINews] High Taste Testers Ep. 1
  article2026-10-17

  Training; Zephon resumed runs on 8, 32 or 64 GPUs from 16-GPU checkpoint with SHA-256-identical global batches

• ClusterMAX 3.0: The Industry Standard GPU Cloud Rating System Returns
  article2026-09-23

  Runs the Llama 3.1 8B pretraining and GPT-OSS MoE training tests.

• Aleph Alpha Kolibri-1 (78B MoE, Apache 2.0)
  repo

  PyTorch-based distributed training infrastructure used to train Kolibri on 768 GPUs with bfloat16 parameters and float32 reductions