← lab
Training internals
One training step, part by part. The step is simulated with real sizes for the model you pick; the memory the rest of the machine holds is live from the GPU box.
Training internalssimulated step · live VRAM from red-gpu-wsl
training internals
Hover any part to see what it does. Run a step with the buttons or keys: 1 forward, 2 backward, 3 optimizer step, then checkpoint (C), crash (X) and resume (R). M switches bf16/fp32, G gradient accumulation, D plays the demo, S stops it.
idle · step 0 · loss – · lr 6.0e-5 · run 414 MB + other 5.2 GB / 24 GB · ckpt noneMade with ISO-GLOW ↗