Fine-tuning time is the number of tokens you train on divided by your throughput in tokens per second. That ranges from minutes for a small LoRA run to days for a full fine-tune of a large model on a large dataset.
The arithmetic
Time is tokens divided by throughput:
- tokens to train on = examples × average tokens per example × epochs
- tokens per second = tokens per second per GPU × GPUs
Example
Take 100,000 conversations of 1,000 tokens each, trained for 2 epochs. That is 200 million tokens.
Halo's docs measure Qwen3-4B, a dense model, at 41,000 tokens per second per GPU. The conditions: one B300 GPU, batch 16 × 2048, no gradient checkpointing, bf16 precision. At that rate:
200,000,000 / 41,000 ≈ 4,900 seconds, or about 81 minutes.
Add the time to load the model, run evaluation and save checkpoints.
What changes the number
- Model size. A larger model processes fewer tokens per second.
- Sequence length. Long rows need more memory each, which forces a smaller batch. Short rows waste time on padding unless you use sequence packing.
- GPUs. Total throughput is the per-GPU figure times the GPU count, but sharding a model across GPUs lowers the per-GPU figure.
- Gradient checkpointing. It saves memory and slows every step.
In Halo
Set enable_efficiency_metrics: true to get tokens per second per GPU in your own logs, then redo the arithmetic with your number. The performance page has the measured table.
