A panel discussion from Microsoft Build featuring experts from Fireworks AI, Unsloth, and Core Automation covering the full stack of custom model deployment — from RL post-training to inference optimization. Key topics include: how GRPO/RL training differs from supervised fine-tuning and why it strains memory (dual inference+training engines), LoRA as a memory-efficient fine-tuning approach (even rank-1 works surprisingly well), gradient checkpointing mechanics, GPU kernel optimization techniques (fusion, CUDA graphs, torch.compile flags), speculative decoding trade-offs for agentic workloads, reward function design challenges and reward hacking, and the emerging role of AI in automating GPU kernel generation. Practical advice includes reading the torch inductor config file, doing math/algorithms before kernel work, and understanding that 80% of managed inference value comes from fusion plus CUDA graphs.