ForgeLM ● live
train LLMs on one RTX 4090
Everyone says you need a datacenter to train a language model. ForgeLM is my answer to that: a complete training stack that fits on a single RTX 4090.
It takes you from raw text to a trained model — tokenizing, training, checkpointing, and evaluating, all driven from a web UI. The point isn't to beat GPT; it's to actually understand what's happening inside, byte by byte, on hardware I can reach out and touch.
// highlights
- From-scratch training: 30M → 500M parameter models
- Runs entirely on one consumer GPU
- Web UI for runs, checkpoints, and evaluation
- Full pipeline: tokenize → train → checkpoint → eval
// stack
- PyTorch
- CUDA
- Web UI