RTX 5090 users optimize Qwen3.5-27B for JSON

The RTX 5090's 32GB VRAM establishes a new standard for 27B-30B models, though aggressive vLLM pre-allocation often triggers premature memory warnings. Blackwell architecture favors FP8 and AWQ over GGUF, and Qwen3.5's hybrid attention allows for massive context windows, up to 128k, when using FP8 KV cache storage and chunked prefill to manage overhead.