NeuralForge: A Queueing-Theoretic Evaluation of Ephemeral Container Isolation for Distributed YOLO Hyperparameter Optimization

Hyperparameter optimization (HPO) for object detection models demands thousands of parallel GPU trials, yet existing distributed frameworks introduce either significant containerization overhead or insufficient state isolation across trials. We present NeuralForge, an open-source MLOps framework that pairs a Celery-based task queue with ephemeral Docker containers to isolate Optuna HPO trials on GPU worker nodes. The system routes trials through prioritized GPU queues (gpus_high, gpus_medium, gpus_low) while enforcing strict memory limits (shm_size >= 16 GB) per container, preventing Out-of-Memory (OOM) failures from propagating to host processes. We evaluate NeuralForge through two complementary methods: (1) a micro-benchmark simulation of N = 1000 dispatch events modeling task arrival patterns on a 3-node RTX 3060 cluster, and (2) an analytical M/M/c queueing model that projects scalability to 30 nodes. The micro-benchmark simulation yields a median task dispatch latency of 0.799 ms (IQR 0.07 ms), a 40.1% reduction in idle GPU time (95% Bootstrap CI [39.88, 40.23]), and a simulated best mAP@50-95 convergence of 0.8209 on COCO with YOLOv8n (95% CI [0.8189, 0.8228]). The M/M/c model projects linear throughput scaling from 1 to 30 nodes with a theoretical maximum concurrent trial capacity of 30. Ablation experiments on Docker memory limits confirm that enforced mem_limit=11g prevents host OOM crashes (median survival >72 h vs. 4.14 h without limits, N=5 seeds). All simulation scripts and configuration artifacts are published for reproducibility.