Trajectory’s concurrent multi‑LoRA stack tackles the core pain points that slow down continual learning workflows. Traditional RL pipelines suffer from long cold‑start delays, often exceeding thirty minutes per run because each job must reload checkpoints, spin up distributed runtimes and warm inference engines from scratch. They also demand massive memory footprints—frontier models can require eight H200 nodes—while most training infrastructure remains single‑tenant, running one experiment at a time and leaving trainers and inference engines idle while waiting on each other. The result is low GPU utilization and sluggish iteration cycles that force teams to wait months for a new model version, risking either disappointing performance or catastrophic behavior for users.
Trajectory addresses these inefficiencies with Continuous Multi‑LoRA Training (C‑LoRA). By keeping a warm, multi‑tenant engine always hot and mapping each experiment to a dedicated LoRA adapter, the system eliminates cold‑start overhead. LoRA adapters freeze the base model and train only small adapter slices, cutting memory usage by an order of magnitude and allowing many experiments to share the same GPU resources. Inference gains come from vLLM’s SGMV decode kernel, which fuses per‑adapter matrix‑vector work into a single GPU launch per decode step, enabling tokens from different adapters to be mixed in the same batch. While training remains single‑adapter at any moment, the scheduler swaps adapter states in and out without freezing, so other tenants continue decoding uninterrupted.
Experimental results on a single H200 node with Qwen3‑4B‑Instruct‑2507 show a 2.81× end‑to‑end experiment‑throughput improvement at eight concurrent runs, with no regression in reward accuracy. Mean experiment time peaks at 1.88× speedup with four concurrent jobs, and all configurations achieve >90% accuracy by step 9. The trade‑off is higher per‑step latency as concurrency grows, but the ideal operating point (N≈2) adds only ~15% rollout time while delivering substantial throughput gains.
For teams building continual‑learning agents—coding assistants, support bots, or any system that must evolve from live feedback—this approach turns training from a serialized bottleneck into a parallel, resource‑efficient process. The full implementation is open source, inviting the community to adapt and scale the method to larger models and more complex tasks.
#AI #Product #MachineLearning #RL #LoRA #ContinualLearning