AWS walks through training Qwen3 three ways on Trainium with torchtitan
AWS ML Blog · rss · 2026-10-22
AWS ML Blog publishes a full walkthrough for training Qwen3-4B on a single trn2.48xlarge (16 Trainium2 accelerators) with torchtitan-neuron, three ways: (1) Transformers v5 as a modeling backend via AutoModelForCausalLM with direct HF safetensors load/save — the verified fine-tune starts at loss 6.54 from published weights vs 12.43 from random init; (2) torchtitan's native Qwen3 implementation with SDPA; (3) NKI mixture-of-experts kernels attached via the override registry. All three share the same torchrun launch pattern and config structure, with pinned versions (torch 2.14.x, transformers 5.16.0, torchtitan-neuron release-2.33), install commands, and full config code included.
More from Infra
- Citi projects Big Four hyperscaler capex to hit $1.62T annually by 2030, $5.35T cumulative — Beth_Kindig · 2026-10-22
- NVIDIA open-sources SimReady Central: 2,000+ sim-ready 3D assets on Hugging Face — lhoestq · 2026-10-22
- From devrel to CTO: juggling 12 cloud and AI providers and still running out of compute — NERDDISCO · 2026-10-22
- Coinbase's x402 was built to kill API keys; Stripe ships Machine Payments SDK for agent payments — kleffew94 · 2026-10-22
- AI software is only ~5% of enterprise IT budgets, says analyst Ben Bajarin — and compute is still constrained — BenBajarin · 2026-10-22
- Invest an hour in Tailscale — it might change your life — EXM7777 · 2026-10-22