Setting Up NanoGPT on AMD MI300X | Generated by AI

Home 2026.06

Here’s what we did:

  1. SSH’d into root@129.212.187.36 (AMD MI300X, Ubuntu)

  2. Installed HF CLI
    • pip3 install -U --break-system-packages 'huggingface_hub[cli]'
    • huggingface_hub 1.17.0
  3. Downloaded lzwjava/zz from HuggingFace
    • hf download lzwjava/zz --local-dir /root/zz
    • 37 files, 424K (non-git download)
  4. Generated SSH key for GitHub
    • ed25519, label lzwjava-amd-devcloud
    • Added to GitHub
  5. Cloned nanoGPT with full history
    • git clone git@github.com:lzwjava/nanoGPT.git
    • 255 commits at /root/nanoGPT
  6. Recloned zz as proper git repo from HF
    • Deleted non-git hf download version
    • git clone https://huggingface.co/lzwjava/zz
    • Set up remotes: origin → GitHub (archived), hf → HuggingFace (active)
    • 18 commits, full history

  7. Opened Zed via zed ssh://root@129.212.187.36//root/zz

  8. Smoke tested download script
    • plan_and_download_fineweb_gpt3_amd.py — works, 267 MB/s
  9. User ran full download in tmux
    • 18 shards, 39 GB, ~10.4B tokens of FineWeb-Edu
    • Done
  10. Installed tokenize deps
    • tiktoken, numpy, pyarrow, tqdm
  11. Smoke tested tokenize script
    • tokenize_fineweb_gpt3.py — works, ~30-50M tok/s
    • Output: uint16, GPT-2 BPE, EOT prefix ✓

Current state: /root/nanoGPT — nanoGPT repo (255 commits) /root/zz — dataset repo (18 commits, 39GB parquet data) Next: — tokenize into .npy shards, then train

Back Donate