Define the training method
Full fine-tuning, LoRA and QLoRA have different memory requirements. Specify your base model, sequence length, optimizer, dataset format and target batch size. Adapter training typically reduces trainable state, but activations and framework overhead still need memory.
Check the complete memory budget
Include model weights, activations, gradients, optimizer state and temporary buffers. Gradient checkpointing, accumulation and quantization change the tradeoffs. Parameter count alone cannot confirm that a run will fit; validate with the exact framework and configuration.
Choose compatible hardware
DGX Spark offers unified memory on an Arm platform, so check Arm-compatible packages and containers; it has no RAID. A100 is a starting point for memory-efficient fine-tuning. V100S uses Volta and lacks native FP8 support; its two GPUs do not automatically form a single memory pool. L40S is also worth evaluating when serving the resulting model is a priority.
Plan data, recovery and scale
Size storage for datasets and checkpoints, and discuss backup options. RAID is not a backup. Larger models or full fine-tuning may need a custom multi-GPU build. Share your model, training method and software versions with an engineer before ordering.
Indicative guidance, not a fit or performance guarantee. Context length, batch size, precision, quantization, KV cache, framework overhead and training method affect memory needs. Validate your exact model and workload with an engineer.
Discuss your workload with a GPU engineer