LLM VRAM & Time Estimator\nPlan memory & runtime for full finetune, LoRA, or RL (GRPO).

Training Mode
Compute Precision (weights/grads)
1 256
1 256
128 16384
1 1024
1 1024
Optimizer
0.3 1
0.5 4
1 256
DeepSpeed ZeRO Stage
0 0.5
0 0.5
Reference Model Precision
1 8192
1 64
0 20
0 40
0.1 200
1000 500000
0 100

VRAM Breakdown (after sharding + overhead)

Per-layer LoRA Params