InferQuest
← Quest map
Training Phase 2 · Efficiency

Go Faster

The modded-nanoGPT lineage: 45 minutes to 74 seconds, one trick at a time.

0/6 · 590 XP
You’re previewing the curriculum. Sign in (top right) to track progress, take knowledge checks, and unlock the verifiers.
Briefing

The NanoGPT speedrun is this path's Boehm worklog: 89 records that turned 45 minutes of GPT-2 training into 74 seconds, each one a named, measured technique — and its biggest export, Muon, went from speedrun trick to torch.optim to training Kimi K2 at a trillion parameters. Read the record history as a method, then earn the ideas by A/B-ing Muon against your own AdamW: the speedrun's discipline (one change, measured, statistically defended) is the actual curriculum.

The skeptic's paper is assigned right next to the hype on purpose: under fair tuning, most claimed 2× optimizer speedups shrink to ~1.1× at even modest scale — hold both facts at once. The capstone milestone is the same discipline pointed at your own loop: make it measurably faster without losing loss, graded as a same-device A/B so your hardware doesn't matter.

Tasks
  • All 89 records and what each changed: Muon, QK-norm, ReLU², untied embeddings, FP8 matmuls, window schedules. The method matters more than any single trick.

    read+80 XPresource ↗
  • Newton-Schulz orthogonalization on the momentum of 2D weights; embeddings/head stay on AdamW. It ships in PyTorch 2.13 as torch.optim.Muon — implement it first, then check yours against the real one.

    build+120 XPresource ↗
  • Fused RMSNorm/RoPE/CE Triton kernels (~20% throughput, big memory cuts) and FlashAttention-2 document-masked packing. Measure, don't trust the README.

    bench+80 XPresource ↗
  • float8 matmuls compose with torch.compile and FSDP2 (1.3–1.5× on real pretraining). On consumer GPUs support varies — knowing where it does and doesn't apply is the point.

    bench+60 XPresource ↗
  • The reality check: under fair tuning, matrix-preconditioner gains shrink toward 1.1× as models grow. Speedrun claims need this next to them.

    paper+50 XPresource ↗
  • The harness trains its own plain fp32 AdamW baseline on a fixed 1M-token workload, then times YOUR loop on an identical model and budget — same device, so any GPU can pass. ≥1.5× wall-clock with final val loss within 0.05 nats. compile, bf16, fused optimizers, Muon — earn it however you like.

    build+200 XPauto-verified