Planner — LLM Toolkit

STEP 1 · WHAT ARE YOU TRYING TO DO?

Default is SFT. LoRA/QLoRA need far less memory than full training. Compare methods →

STEP 2 · MODEL

Model size (B parameters)

Billions of parameters, e.g. 1 = 1B.

Advanced architecture (layers, hidden size, heads)

Layers

Hidden size

Must be divisible by attention heads.

Attention heads

STEP 3 · DATA

Token budget (tokens)

Total training tokens. Calculate →

Sequence length (tokens)

Need a pipeline? Open Data Planner →

STEP 4 · HARDWARE

GPU VRAM (GB per GPU)

Check your GPU →

GPU count

Precision

BF16/FP16 use 2 bytes/param; FP32 uses 4.

STEP 5 · TRAINING SETTINGS

Optimizer

Batch size (per GPU)

Gradient accumulation steps

Advanced training (learning rate, scheduler, checkpointing)

Learning rate

Scheduler

Gradient checkpointing (trades ~10× less activation memory for ~20% slower steps)

What the status means

RECOMMENDED fits with >20% headroom · POSSIBLE fits with 5–20% headroom · TIGHT fits with <5% headroom, may OOM · OOM exceeds VRAM. All values are estimates, not measured guarantees — verify on hardware.

Training Planner