dgx-lab-benchmarks-vs-reality-day-4

After 6 days of intensive testing on NVIDIA's DGX Spark, training performance matches published benchmarks (80+ tokens/sec), but critical production issues emerge: GPU inference completely fails with standard PyTorch workflows, memory fragmentation forces 2-3 hour training session limits, and llama.cpp direct invocation produces empty responses. The ARM64 + Blackwell + CUDA 13.0 combination requires expert-level workarounds including manual cache clearing every 50 steps, CPU-based inference for evaluation, and Ollama for stable production inference. Successfully trained 7 models with 70-84% accuracy using documented mitigations.