SLM Fine-Tuning Planner

Estimate time, memory fit, and budget-aware infrastructure options for fine-tuning small language models.

1. Model

Model Input Mode

Fetch from Hugging Face Enter Manually

Hugging Face model name

Examples

Waiting for a model lookup.

Paste a Hugging Face model id and fetch metadata, or switch to manual mode.

Workload input basis

Use examples when you know dataset rows. Use total tokens when you already know the token volume per epoch.

Examples / sequences Total tokens

Dataset size

Epochs

Sequence length

Training mode

This changes both memory and throughput assumptions.

Precision

4bit is intended for QLoRA-style adapter training. Use fp16 or bf16 for fuller training flows.

Assume gradient checkpointing

Enabled checkpointing lowers memory at the cost of some extra compute.

Number of GPUs to simulate

Comparison rows use this GPU count. Budget optimization searches 1, 2, 4, and 8 up to this limit.

1 8

GPUs to compare

T4 L4 A100 H100 RTX 4090

Clouds to include in budget recommendation

AWS Azure GCP

Cloud pricing mode

on_demand spot

Budget ($)

Use measured throughput calibration

Recommended when you already ran a short pilot fine-tuning job.

Configure the workload and run an estimate.

Results will appear here.

No run yet.

Run the planner to see the active assumptions.

Training Time Comparison