I'd say the first step is to rule out LoRA. If LoRA is an option, it buys you more than just the rig savings: - You can deploy multiple specialized LoRAs for d

I'd say the first step is to rule out LoRA. If LoRA is an option, it buys you more than just the rig savings: - You can deploy multiple specialized LoRAs for different tasks - It massively reduces your train-test latency - You get upstream LLM updates for "free," maybe you can even add the training to your CI