TX · Utils
Utilities
Small interactive tools I've generated for my own use.
-
Mist Box LLM Training Reference
Which open LMs fit pretrain / full fine-tune vs LoRA / QLoRA on an RTX 3070 8 GB (BF16, training overhead included). Target: Mist box — 94 GB RAM, 16 cores.
mist.train.fit() -
Transformer GPU Memory Calculator
GPT and GPT-NeoX / Pythia decoder presets with architecture sizes, parameter counts, and BF16 weight memory.
model.size() -
LLM Training Data Calculator
If your model has N parameters, how much text does it need? Quick estimates using Chinchilla-optimal and legacy GPT-2 era scaling ratios.
data.tokens(N) -
Google Colab Accelerator Reference
Peak memory and FLOPs for an RTX 3070 vs. Google Colab accelerators, with relative speed vs. the local baseline.
accel.ref() -
Accelerator Cost Comparison
GPUs available on 3+ of Colab, Vast, Lambda, Modal, and RunPod — plus Colab TPUs — with hourly cost side by side. Includes Lambda, Modal, and RunPod catalogs.
accel.cloud($/hr) -
Vast.ai GPU Accelerator Reference
Vast marketplace GPUs with release year, peak FLOPs, on-demand $/hr, hours from $50 and $100, and speed vs. an RTX 3070 baseline. Sort by year, FLOPs, price, or RAM.
vast.gpu.ref()