Skip to content
View pjordanandrsn's full-sized avatar
3 followers · 7 following

Achievements

Achievement: Pull Sharkx4Achievement: QuickdrawAchievement: Pair Extraordinairex4

Achievements

Achievement: Pull Sharkx4Achievement: QuickdrawAchievement: Pair Extraordinairex4

Highlights

  • Pro

Block or report pjordanandrsn

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
pjordanandrsn/README.md

Jordan Anderson

I make large models run on smaller machines.

Large models. Smaller machines. Plan, run, measure, and keep the evidence.

Research · Work with me · Hugging Face

I build open-source ML systems, GPU kernels, and fixes for the tools they depend on.

Start with Loggetta

pip install loggetta
loggetta inspect
loggetta plan Qwen/Qwen3-30B-A3B --seq 2048

One install. Check the machine, plan a supported training run, and keep the results. New in Loggetta 0.6.0: an AMD GPU is found and refused by name, and with experts4bit-qlora 0.53.0 or later the grouped_nf4 MoE backward is priced once.

Project What it does
Loggetta Plans MoE training for your hardware and records what ran.
experts4bit-qlora Fine-tunes and serves large MoEs using GPU memory, RAM, and SSD storage.
grouped-nf4-gemm Runs expert math directly on packed 4-bit weights.

Results worth a look

  • Unsloth spends 1.92× e4b's GPU time per training step on Qwen3-30B-A3B / RTX 5090, and 2.80× its wall-clock time on an AMD EPYC 7713 host: matched work and software, comparable held-out loss. Unsloth peaked lower (24.27 vs 26.16 GB). Benchmark. Earlier wall-clock readings, before e4b's current defaults: 2.352× and 2.468×.
  • 120B QLoRA experiment at 9.82 GB peak VRAM, using host memory. All 144 frozen expert tensors stayed byte-identical after training. Native-MXFP4 experiment.
  • Long 4,096-token rows: Unsloth spends 1.160× e4b's GPU time per step on an RTX 5090; the wall-clock ratio varies with the host (1.287–1.773×). Latest result and limits.

I publish the hardware, code, controls, and failed tests alongside the wins. Evidence and current status

Fixes merged upstream

Project Contribution
OpenVINO Intel GPU compilation fixes and regression coverage: #35661, #35712, #36017.
bitsandbytes / Axolotl Stop checkpointed training from retaining dequantized parameters in the cache: #1999, #3915.
Transformers / Unsloth Zoo Fix end-of-sequence handling and MoE weight preprocessing: #47931, #1082.
Unsloth Zoo Add selective response-token loss masking (including </think>) across text, vision-collator, and MLX paths, with regression tests: #852.
zkbugs Add the public ZisK signed-remainder soundness bug record: #97.

Also: JWST image research, security disclosures, and technical and policy consulting.

Contact and public GPG key

Email · Download public key · Verify on my website

Ed25519 · expires 11 June 2028.

DAEF C03B 9188 793C 80E5 20C8 68F0 6663 C7FB 250F

Pinned Loading

  1. loggetta loggetta Public

    Plan and run Mixture-of-Experts fine-tuning on one GPU: check the machine, fit the model, and keep a report of each run.

    Python 1

  2. experts4bit-qlora experts4bit-qlora Public

    Fine-tune and serve large Mixture-of-Experts models on one GPU, with 4-bit experts in GPU memory, RAM or SSD.

    Python 3

  3. grouped-nf4-gemm grouped-nf4-gemm Public

    4-bit GPU kernels that compute directly on packed Mixture-of-Experts weights, for training and inference.

    Python 2

  4. arr-malware-screen arr-malware-screen Public

    Malware screen for *arr download queues — two-tier regex over realized torrent file lists, delete + blocklist via the *arr API. qBittorrent/Deluge/Transmission.

    Shell