We've been thinking a lot about inference infrastructure recently, and it seems like the challenges are very different from training. Training tends to be compute-heavy but predictable, while inference introduces things like: - latency constraints - dynamic batching - unpredictable traffic patterns - model versioning in production - GPU utilization issues at smaller batch sizes For people running inference in production today, what's been the most painful part?