15-year-old builds from-scratch SLM trainer in Python+Rust, seeks GPU testers
PainterNo8889 · reddit · 2026-11-14
A 15-year-old developer is building SLM-CREATOR, a from-scratch, hardware-aware trainer for small text models (capped at 100M params) in Python and Rust — hand-coding GPU detection, sizing, a BPE tokenizer and the training loop. Development happens on a CPU-only machine, so they're asking NVIDIA/AMD (incl. ROCm) users to run architecture.py on real hardware. Technical details: vendor CLI parsing (nvidia-smi/amd-smi/rocm-smi) with PCIBUSID ordering to map multi-GPU devices correctly; a TFLOPS benchmark with a 12-second thermal soak and median of 10×20 matmuls at 4096×4096, falling back to fp32 when bf16 is software-emulated; and a Rust sizer computing max params from VRAM, training hours, TFLOPS and Chinchilla/data limits. Testing takes 2 minutes; feedback goes to Issue #1.
More from Infra
- Cloud Reports Measure the Wrong Thing: Hybrid Cloud Is Often Legacy, Not Strategy — DavidLinthicum · 2026-11-14
- Will Depue claps back at Teortaxes over claim that 2-3K Blackwell runs aren't "data-center scale" — teortaxesTex · 2026-11-14
- Local LLM stack cheat sheet: seven open models mapped to chat, coding, TTS and more — alex_verem · 2026-11-14
- Qwen3.8-Flash NVFP4 wows on V100s: insanely fast prefill with Strata — Scobleizer · 2026-11-14
- How a $1M 'fake chip' saved Nvidia in 1997 — the RIVA 128 bet that made Jensen Huang — bigaiguy · 2026-11-14
- SGLang Summit 2026 swag pins turn insider jokes like CUDA OOM and KV Cash into collectibles — ying11231 · 2026-11-14