Loggetta
Inspect, plan, run, and save reports.
Jordan Anderson · ML systems & GPU kernels
Large models.
Smaller machines.
I build open-source tools to fine-tune large models on a single GPU. Check the machine, plan the run, and keep the results.
One install. Three projects.
pip install loggetta
loggetta inspect
loggetta plan Qwen/Qwen3-30B-A3B --seq 2048experts4bit-qlora
Load, fine-tune, serve, and offload model weights.
grouped-nf4-gemm
Compute on packed 4-bit weights and move expert data.
New in Loggetta 0.4.0: MoE training estimates price the grouped_nf4 backward pass, and resident grouped_nf4 training runs again with experts4bit-qlora 0.49.0 or later.
Results worth a look
1.92×
Unsloth's GPU time per training step, against e4b's
Qwen3-30B-A3B / RTX 5090. Matched work and software; comparable held-out loss. Wall-clock 2.80× on an AMD EPYC 7713 host. Unsloth peaked lower (24.27 vs 26.16 GB).
Read the benchmark →9.82 GB
120B QLoRA experiment
Peak GPU memory, using host RAM. All 144 frozen expert tensors stayed byte-identical. A native-MXFP4 experiment, separate from Loggetta’s training path.
Read the experiment →1.160×
Unsloth's GPU time on 4,096-token rows, against e4b's
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 →Speed results use torch 2.12.1+cu130 / transformers 5.5.0 and apply to the named workloads.
Fixes merged upstream
GPU compilation in OpenVINO. Training-cache fixes in bitsandbytes and Axolotl. End-of-sequence handling in Transformers. MoE preprocessing in Unsloth Zoo.
I publish the controls and failed experiments alongside the wins. Research notes · Results as JSON
Contact and public GPG key
Email · Download public key · Verify fingerprint
Ed25519 · expires 11 June 2028DAEF C03B 9188 793C 80E5 20C8 68F0 6663 C7FB 250F