Open-source trainer beats ms-swift by 58%, hits 12,523 tokens/s on one GPU
incorporo · reddit · 2026-11-03
A Reddit user open-sourced training optimizations for BH200/H100/A100 datacenter GPUs covering both text-to-image (IdeogramV4-class) and LLM (Qwen3.5-9B-class) trainers.
Key results:
- 58% faster than ms-swift, 76% faster than Unsloth, and 178% faster than transformers + PEFT on identical setups;
- 12,523 tokens/s training Qwen3.5-9B at 53.9% clock-adjusted MFU (power-capped by the test machine); baselines: ms-swift 24%, Unsloth 22%, transformers+PEFT 14%;
- Loss parity with <6e-4 difference; micro-batch sweep found a 43.7% best-config headroom.
Step-time ablations are included; code is at github.com/logolabs/perfkit.
More from Infra
- Periodic Labs to detail custom RL infra for physical experiments at SGLang Summit 2026 — hsu_byron · 2026-11-03
- NVIDIA Research Blog: Why KV Cache Compression Breaks in Production vLLM Deployments — blaizedsouza · 2026-11-03
- mlx-learn launches: scikit-learn API on Apple silicon with up to 30X training and 300X inference speedups — HankYeomans · 2026-11-03
- Podcast episode: Is running local AI for you? — rez0__ · 2026-11-03
- ByteDance crawler spotted scraping websites at massive scale — gefei55 · 2026-11-03
- 1e29 FLOPS would take 200k GB200s 3.2 years and 6.2 TWh, dev notes — mike64_t · 2026-11-03